The '🤖machines will replace doctors👩🏽⚕️' debate turns 70! From Warner Slack to Geoffrey Hinton. (Part 1 of 5)
What do the invention of television and fuzzy logic have to do with AI in medicine? More than you'd think.
Welcome to AI Health Uncut, a brutally honest newsletter on AI, innovation, and the state of the healthcare market. If you’d like to sign up to receive issues over email, you can do so here.
Fun fact: The same person who invented television also played a significant role in the early development of AI in medicine.
Another fun fact: PROMIS, the first medical record system I discuss in this article, was already using a touch-screen terminal back in 1967!
History isn’t just fascinating. It’s full of crossovers that make Marvel jealous. 😊
In this 4-part series, I thoroughly review the 70-year history and research on the ‘machines replacing doctors’ debate.
But first, some housekeeping:
First, apologies to my paid subscribers and Founding Members who’ve been patiently waiting for Part 2 of my series, “These 14 Health AI Companies Have Been Lying About What Their AI Can Do.” It’s coming, and trust me, it’s going to be a bombshell. In the meantime, here’s a little (and super obvious 😊) pre-release quiz: Which so-called health AI startup just raised $165 million, is valued at $0.5 billion, and is riding on stolen tech? Stay tuned for Part 2...
I’m looking for people who are worried about the future of healthcare but hesitant to speak up, whether out of fear of retaliation from their boss or other concerns. I’m particularly interested in compelling inside stories from digital health and health AI. I believe that combining “mosaic theory” analysis with firsthand accounts from within companies is the key to revealing both the positives and the negatives in our industry. This approach is how my recent pieces on Walmart Health and IBM Watson Health came together—thanks to individuals who stepped forward to share their stories. Friends and colleagues who’ve known me for years can vouch for my discretion and integrity. I will always honor anonymity, and I will never act without a person’s consent.
Now, for the big announcement:
🚨 Attention, my paid subscribers and Founding Members: In 2025, I’ll be launching a new series of articles, tentatively titled “Venture Capital is a Scam: Building a Better Way to Fund Innovation,” with a sharp focus on the digital health sector. I want to expose the net harm that venture capital has inflicted on digital health innovation over the past 20 years. Innovation in digital health is at an all-time low, and it breaks my heart. This is an urgent matter. The goal of this research is to explore alternative methods of financing and capital allocation in digital health and healthcare AI. VCs are optimizing their picks and managing startups in their self-interest—not necessarily in the interest of founders, and certainly not in the interests of customers, patients, and clinicians.
With this in mind:
I welcome any ideas, research, and inside stories from within the VC world. A significant portion of what I write is heavily based on these insights. I understand that trust and loyalty are everything, and I will never disclose anything without your consent.
The amount of information and data I’ve collected so far could fill, if not a whole book, then at least half of one. I know it’s a long shot, but if anyone is interested in financing my research in the form of a book 📚, I believe it could greatly benefit the healthcare industry. This book could help break the healthcare oligopoly, expose the shady schemes of venture capital, and create a better way to build an ecosystem of innovation in healthcare. I’m a one-man operation—no sponsors, no advertisers—and unfortunately, I don’t currently have the means to write a book like this on my own.
Thank you!
All right, let’s get back to our story.
First things first—let’s all take a deep breath and lay out the facts.
The debate over who’s better and when, exactly, the medical profession will fade into the abyss has reared its head once again last week. So I’m here to calmly summarize the latest research on the subject. Yes, there will be some sarcasm sprinkled in, but the goal is simple: gather the research, opinions, and quotes from people much smarter than me (all 184 of them 😊), and let the readers make up their own minds.
Following the lively discussion on X spaces organized by Sanat Dixit, MD, FACS last week—where I had the honor of being a guest speaker—there was a robust exchange of ideas from notable medical professionals such as Anthony DiGiorgio, DO, MHA, Owen Scott Muir, MD, DFAACAP, Anish Koka, MD, Chuck-G, John Paul G. Kolcun, MD, Krishnan Chittur, and many other experts. This led to a fresh debate on whether AI will be “replacing” doctors anytime soon. We covered a lot of ground during that conversation, and the recording is available here.
And now, here are all the honorable mentions in this article—yes, all 299 of them. 😊
Part 1:
Walmart Health, Sanat Dixit, MD, FACS, Anthony DiGiorgio, DO, MHA, Owen Scott Muir, MD, DFAACAP, Anish Koka, MD, Chuck-G, John Paul G. Kolcun, MD, Krishnan Chittur, Warner V. Slack, MD, Center for Clinical Computing, Harvard Medical School, Beth Israel Deaconess Medical Center, ELIZA, MIT, Joseph Weizenbaum, Natural Language Processing (NLP), Turing Test, Rogerian psychotherapist model, F.A. Nash, Cornell Medical School, Mt. Sinai Hospital, IBM, Greene JA, B.J. Davis, Lipkin M, Keeve Brodman, van Woerkom AJ, Erdmann AJ Jr, Goldstein LS, Andrew S. Lea, Max Planck Institute for the History of Science, Weill Cornell Medicine Library, Robert Ledley, Lee Lusted, Georgetown, University of Rochester, Society for Medical Decision Making, Warner HR, Toronto AF, Veasy LG, de Dombal FT, Leaper DJ, Staniland JR, McCann AP, Horrocks JC, Robert W. Taylor, DARPA, ARPA, XAI (Explainable Artificial Intelligence), Vladimir Zworykin, Boris Rosing, Paul Langevin, Vladimir Lenin, Admiral Alexander Kolchak, Marie Curie, Bane F, Wander A, Ralph Engle, MD, The Rockefeller Institute for Medical Research, McTernan E, Crocker D, Mullainathan S, Obermeyer Z, Bruce G. Buchanan, Edward H. Shortliffe, Robert K. Lindsay, Edward A. Feigenbaum, Joshua Lederberg, Schenthal JE, Sweeney JW, Nettleton WJ, Yoder RD, Vandenberg, Weinrauch, Hetherington, Judy Faulkner, Epic Systems, Problem-Oriented Medical Information System (PROMIS), Jan Schultz, Lawrence L. Weed, M.D., Cleveland Metropolitan General Hospital (CMGH), University of Vermont’s Medical Center Hospital of Vermont (MCHV), Digiscribe, Control Data Corporation, FORTRAN, SNOBOL, V77-400 Varian Data Machines minicomputers, Xerox Palo Alto Research Center (PARC), Michael F. Collen, American Medical Informatics Association, Clement J. McDonald, Regenstrief Institute, National Library of Medicine Archives, John Rees, Casimir A. Kulikowski, Sholom M. Weiss, Rutgers University, P. Szolovits, Clancey W. J., Juri Yanase, Evangelos Triantaphyllou, R. Hirani, K. Noruzi, H. Khuram, A. S. Hussaini, E. I. Aifuwa, K. E. Ely, J. M. Lewis, A. E. Gabr, A. Smiley, R. K. Tiwari, M. Etienne, Dr. Jack D. Myers, Dr. Harry E. Pople Jr., Dr. Randolph A. Miller, F. E. Masarie, Klaus-Peter Adlassnig, Lotfi A. Zadeh, Masaki, Watanabe, Bell Labs, G. Kolarz, W. Scheithauer, H. Effenberger, G. Grabner, P. Klinov, B. Parsia, D. Picado-Muiño, Dr. G. Octo Barnett, J. J. Cimino, J. A. Hupp, E. P. Hoffer, M. J. Feldman, G. Elhanan, S. A. Socratous, S. P. Bartold, G. G. Hannigan, Eta S. Berner, Tonya J. La Lande, CASNET, INTERNIST-I, MYCIN, DENDRAL, CADUCEUS, QMR, CADIAG, DXplain.
Part 2:
Warner V. Slack, MD, Vinod Khosla, Geoffrey Hinton, Ezekiel J. Emanuel, MD, Ph.D., Phelps Kelley, MD, Yuval Noah Harari, Ali Parsa, Judy Faulkner, Ben Horowitz, Derya Unutmaz, MD, Bindu Reddy, Elon Musk, Eric Topol, Statista, National Institutes of Health (NIH), Tsarnick, Lauren Silverman, National Public Radio (NPR), Greg Ip, The Wall Street Journal, Enrico Coiera, Epic Systems, Ann Richardson, Marc Andreessen, Ayers JW, Poliak A, Dredze M, JAMA Internal Medicine, Jonathan Reisman, MD, The New York Times, Med-PaLM 2, Singhal K., Azizi S., Tu T., Mahdavi S.S., Wei J., Chung H.W., Matias Y., NPJ Digital Medicine, Esteva A., Robicquet A., Ramsundar B., Kuleshov V., DePristo M., Chou K., Dean J., Nature Medicine, Beger J., Newsweek, Harvard Health Publishing, Jiang F., Jiang Y., Zhi H., Dong Y., Li H., Ma S., Wang Y., Stroke and Vascular Neurology, World Health Organization, HealthSpot, Theranos, Babylon Health, CarePod, Forward Health, Hippocratic AI, MayaMD, Versel N., MedCity News, Elizabeth Holmes, Ramesh “Sunny” Balwani, John Carreyrou, Weaver C., Copeland R., Schwartz B., TechCrunch, Alexander P., Lomas N.
Part 3:
Yann LeCun, Ph.D., Curtis Langlotz, MD, Ph.D., David D. Luxton, Ph.D., MS, Nirav R. Shah, MD, MPH, Geraint Rees, Ph.D., Antonio Di Ieva, MD, Ph.D., Eric Topol, MD, Andrew Ng, Sanat Dixit, MD, FACS, Guy Culpepper, MD, Mark Cuban, McGill University, Declan O’Regan, Center for Artificial Intelligence in Medicine & Imaging (AIMI), Stanford University, Filippo Pesapane, Marina Codari, Francesco Sardanelli, AMA Journal of Ethics, AlphaZero, Nobel Prize, World Economic Forum, The Lancet, The Doctors Company, Google Brain, DeepLearning.AI, TED Talk, Mark Cuban Cost Plus Drug Company, American Academy of Family Physicians (AAFP), Riedl, D., Schüßler, G., Ann M. Richardson, EPIC Systems, Judy Faulkner, Frontiers in Psychology, Rishad Usmani, MD, Spencer Dorn, FDA (Food and Drug Administration), Obermeyer, Z., Powers, B., Vogeli, C., Mullainathan, S., Annals of Internal Medicine, The Register, JAMA Open, Watson for Oncology, Science (journal), Theranos, HealthSpot, Babylon Health, Forward Health, CarePod, Hippocratic AI, MayMD, Deep View Report, McKinsey, Meerkat 70B, Jiang, F., Sutton, R. T., Froomkin, A. M., Kerr, I., Pineau, J., Nature Medicine.
Part 4:
MYCIN, INTERNIST-I, Electronic Health Record (EHR) systems, IBM, Thomas J. Watson, Jr., William B. Schwartz, M.D., New England Journal of Medicine, Wall Street Journal, iPhones, iPads, Gizmodo, Watson, Columbia University, University of Maryland, Nuance Communications, Inc., Apple, Harvard Medical School, Dr. Warner Slack, Forbes, Watson Health, American Medical Association (AMA), AMA Journal of Ethics, IBM Watson, Luxton DD, November JA, Johns Hopkins University Press, Yu KH, Beam AL, Kohane IS, Nature Biomedical Engineering, Greene JA, Andrew S. Lea, OpenAI o1 Strawberry model, OpenAI, ChatGPT, X-ray, MRI.
Wow, that’s a ridiculously long list—299 unique mentions, to be precise. We’ve got 184 individuals, 79 companies/organizations, and 36 products/systems.
Before we dive in, let me make one thing clear. I’m over this tired line, but I have to mention it for the sake of completeness: “AI won’t replace doctors, but doctors who use AI will replace those who don’t.” It’s been thrown around for at least 5 years now (any guesses on who said it first?). At this point, it’s obvious, and frankly, I’m sick of hearing it.
Now, here’s how we’re going to attack this article:
Part 1:
1. 1955: Warner V. Slack, MD and The Dawn of the ‘Machines Will Replace Doctors’ Debate
1.1. Automated Patient Questionnaire (1955)
1.2. Patient-Centered Computing and Cybermedicine (1965 onwards)
1.3. Center for Clinical Computing
1.4. Association with MIT’s ELIZA
1.5. Views on Soliloquy and Mental Health
1.6. Warner Slack’s Legacy and Impact
2. 1959 and Beyond: Pioneering Machine Learning for Diagnosis—From Nash to Zworykin
3. 1960s: PROMIS, the First Touch-Screen Computer-Based Medical Record System
4. 1960s: CASNET, the First AI System for Medical Diagnosis and Treatment
5. 1970s and 1980s: The Rise of Artificial Intelligence in Medicine through Expert Systems
5.1. INTERNIST-I: The First AI Model for Internal Medicine Diagnostic Reasoning
5.2. MYCIN: Bridging Artificial Intelligence and Medical Diagnosis
5.3. CADUCEUS: A Successor to INTERNIST-I with Comprehensive Knowledge Base
5.4. QMR: A Superior Computational Knowledge Base Outpacing INTERNIST-I and CADUCEUS
5.5. CADIAG: The Birth of Fuzzy Logic in Medical Diagnosis
5.6. DXplain: The Pioneer of Computer-Based Diagnostic Decision Support
Part 2:
6. Argument FOR: AI Will Replace Doctors – It’s More Powerful, More Accurate, and Could Even Be More Empathetic
6.1. Famous Quotes for the “AI Will Replace Doctors” Argument
6.1.1. Warner V. Slack, MD, Late Professor at Harvard Medical School and medical informatics pioneer
6.1.2. Vinod Khosla, “The Forecaster Without the Horizon”, a venture capitalist and co-founder of Sun Microsystems
6.1.3. Geoffrey Hinton, “The Godfather of AI” and freshly minted 2024 Nobel Prize laureate in physics
6.1.4. Phelps Kelley, MD, a diagnostic radiologist
6.1.5. Ezekiel J. Emanuel, MD, Ph.D., prominent physician and architect of the Affordable Care Act (ACA)
6.1.6. Yuval Noah Harari, historian, philosopher, and best-selling author
6.1.7. Ali Parsa, “The Madoff of Digital Health,” CEO of the now-bankrupt and fraudulent Babylon Health
6.1.8. Judy Faulkner, Founder and CEO, Epic Systems
6.1.9. Ben Horowitz, co-founder of Andreessen Horowitz (a16z) venture capital firm
6.1.10. Derya Unutmaz, MD, professor of immunology at the Jackson Laboratory for Genomic Medicine
6.1.11. Bindu Reddy, founder of Abacus.AI
6.1.12. Elon Musk, entrepreneur and founder of Tesla, SpaceX, Neuralink, and The Boring Company
6.2. Advancements in AI Empathy and Patient Interaction
6.2.1. AI Outperforms Doctors in Empathy
6.2.2. AI Enhances Patient Satisfaction
6.3. Superior Diagnostics and Treatment Planning by AI
6.3.1. AI Outperforms Human Doctors in Diagnostics
6.3.2. AI in Personalized Treatment Planning
6.4. AI Reducing Physician Burnout and Improving Efficiency
6.4.1. Automation of Administrative Tasks
6.4.2. Mitigating Physician Shortages
6.5. Economic and Accessibility Benefits
6.5.1. Reducing Healthcare Costs
6.5.2. Improving Accessibility in Underserved Areas
6.6. AI Tools That Tried to Replace Doctors and Nurses
6.6.1. HealthSpot: Telemedicine Kiosks That Missed the Mark
6.6.2. Theranos: The Fallacy of Overpromised Technology
6.6.3. Babylon Health: Lied About Its AI
6.6.4. Forward’s CarePod: A Tech-First Approach to Primary Care
6.6.5. Hippocratic AI: Replacing Nurses with Algorithms
6.6.6. MayaMD: Questionable Claims of AI Outperforming Doctors
Part 3:
7. Argument AGAINST: AI Won’t Replace Doctors—But Every Doctor Will Have an AI Assistant / AI Agent
7.1. Famous Quotes for the “AI Won’t Replace Doctors” Argument
7.1.1. Yann LeCun, Ph.D., Chief AI Scientist at Meta and founding father of convolutional neural networks (CNN)
7.1.2. Declan O’Regan, Professor of of Cardiovascular AI
7.1.3. Curtis Langlotz, MD, Ph.D., Stanford radiologist and AI pioneer
7.1.4. David D. Luxton, Ph.D., MS, an expert in the field of artificial intelligence and health care technology
7.1.5. Nirav R. Shah, MD, MPH, a prominent physician and health policy expert
7.1.6. Geraint Rees, Ph.D., Professor of cognitive neurology, University College London
7.1.7. Antonio Di Ieva, MD, Ph.D., internationally acclaimed neurosurgeon
7.1.8. Eric Topol, MD, a prominent cardiologist and digital medicine expert
7.1.9. Andrew Ng, a prominent data scientist, founder of Google Brain and DeepLearning.AI
7.1.10. Sanat Dixit, MD, FACS, Professor of Neurological Surgery
7.1.11. Mark Cuban, an entrepreneur and a founder of Mark Cuban Cost Plus Drug Company
7.1.12. Guy Culpepper, MD, a nationally recognized primary care physician and Fellow of the American Academy of Family Physicians (AAFP)
7.2. The Irreplaceable Doctor-Patient Relationship
7.2.1. Empathy and Trust in Healthcare
7.2.2. The Therapeutic Power of Human Interaction
7.2.3. Patient Autonomy and Empowerment
7.3. Clinician Skepticism and Resistance
7.3.1. Skepticism Rooted in Experience
7.3.2. The Emotional and Ethical Dimensions of Care
7.4. Best Practices: Legal and Ethical Responsibilities
7.4.1. Best Practices in Medicine Are Inconsistent and Evolving
7.4.2. Ethical and Responsible AI in Medicine Lags Behind
7.4.3. Explainability and Trust are Still Major Concerns
7.4.4. Algorithmic Biases and Health Equity are Far From Solved
7.4.5. Legal and Ethical Concerns: Doctors Go to Jail, Engineers Don’t
7.5. FDA’s Lack AI Oversight Is Failing Medicine
7.5.1. Lack of Robust Regulations for AI Systems
7.5.2. Slow Feedback Loops
7.5.3. Legal Liability and Clinical Decision-Making
7.5.4. AI’s Biases and FDA’s Lack of Response
7.5.5. FDA’s Inadequate Validation Standards
7.5.6. FDA’s Failure to Promote Interdisciplinary Collaboration
7.5.7. The Data Problem
7.5.8. Economic Pressures
7.5.9. Lobbying
7.5.10. The Revolving Door
7.5.11. Misaligned Incentives
7.6. AI in Healthcare Standardization Challenges
7.6.1. Evolving Standards for AI in Medicine
7.6.2. International Regulatory Variations
7.6.3. Data Standardization
7.7. AI Biases, Hallucinations, and Data Limitations
7.7.1. AI Biases
7.7.2. AI Hallucinations
7.7.3. Data Limitations
7.8. Limitations of AI Technology in Healthcare
7.8.1. Lack of Contextual Understanding and Clinical Nuance
7.8.2. Challenges with Autonomous Clinical Decision-Making
7.8.3. Unstructured and Incomplete Data in Healthcare
7.8.4. Limited Generalization Across Medical Disciplines
7.8.5. Complexity of Clinical Judgment and Human Intuition
7.8.6. Technological Readiness and Integration Challenges
7.8.7. Limited Application in Certain Medical Tasks
7.9. Increased Workload and Reduced Efficiency: The EHR Nightmare All Over Again?
7.9.1. Increased Workload: The Unfortunate EHR Parallel
7.9.2. The EHR Nightmare All Over Again?
7.9.3. Reduced Efficiency: The Editing Paradox
7.9.4. AI is Not Yet Ready for Clinical Autonomy
7.9.5. Learning Curve and Fixed Costs
7.9.6. Feedback Loop Deficiencies
7.10. The Complexity of Human Health and the Need for Clinical Judgment
7.10.1. The Complexity of Human Health
7.10.2. Multifactorial Nature of Health Problems
7.10.3. The Need for Clinical Judgment
7.11. Historical Failures in Replacing Doctors with Technology
7.11.1. Theranos: From Blood Work to Con Work
7.11.2. HealthSpot: The Telemedicine Kiosk Failure
7.11.3. Babylon Health: The Emperor Had No Clothes
7.11.4. Forward’s CarePod: False Start
7.11.5. Hippocratic AI: Nurses Under Threat?
7.11.6. MayMD: Repeating the Mistakes of the Past
7.12. Venture Capital: Focusing on Quick Bucks, Not Healthcare
7.12.1. Venture Capital’s Incentive Structure Prioritizes Quick Returns
7.12.2. The Profit Motive Encourages Shortcuts and Misaligned Priorities
7.12.3. Lack of Long-Term Focus Hurts Innovation in Healthcare AI
7.12.4. The Misallocation of Capital in Healthcare AI
7.13. AI as a Tool for Augmentation, Not Replacement
7.13.1. Augmenting, Not Replacing, Critical Thinking
7.13.2. AI Systems Are Not Yet Ready for Autonomous Decision-Making
7.14. The Future: Every Doctor Will Have an AI Assistant / AI Agent
Part 4:
8. 70 Years Later: Machine Learning Evolves, Healthcare Stays the Same
9. Two Essential Conditions for Industry-wide AI Adoption
10. Conclusion
Alright, let’s kick things off with Part 1.
1. 1955: Warner V. Slack, MD and The Dawn of the ‘Machines Will Replace Doctors’ Debate
Warner V. Slack, MD, a pioneer in medical informatics, was a visionary who foresaw the transformative impact of computers on healthcare. His contributions spanned from developing automated patient questionnaires to advocating for patient-centered computing and exploring the therapeutic potential of human-computer interactions. His insights into soliloquy and mental health remain relevant, highlighting the importance of self-reflection and communication in emotional well-being.
1.1. Automated Patient Questionnaire (1955)
Warner V. Slack, MD, began exploring the use of computers in healthcare as early as 1955. He recognized the potential of automated systems to collect patient data efficiently. His work led to the development of one of the first computerized patient questionnaires. This system allowed patients to input their medical histories directly into a computer, streamlining data collection and reducing the burden on healthcare professionals.
1.2. Patient-Centered Computing and Cybermedicine (1965 onwards)
By 1965, Dr. Slack had expanded his focus to patient-centered computing. He was a strong advocate for empowering patients by giving them direct access to medical information and computational tools. This approach laid the groundwork for what would later be known as cybermedicine—the use of internet and digital technologies to deliver healthcare services remotely.
1.3. Center for Clinical Computing
Dr. Slack co-founded the Center for Clinical Computing at Harvard Medical School and Beth Israel Deaconess Medical Center. The center became a hub for innovation in medical informatics, focusing on developing systems that improved patient care through technology. Under his guidance, the center worked on projects that integrated computer systems into clinical practice, enhancing both patient and provider experiences.
1.4. Association with MIT’s ELIZA
ELIZA was a groundbreaking program at MIT developed by Joseph Weizenbaum in the mid-1960s, designed to simulate a conversation between a human and a computer. Eliza was built as an early example of natural language processing (NLP), but it wasn’t initially created with the goal of passing the Turing Test—a benchmark for determining if a machine can exhibit intelligent behavior indistinguishable from a human.
In fact, Eliza was designed as a spoof, meant to satirize how shallow and mechanical computer communication was at the time. Weizenbaum created the program to show how simple pattern-matching could produce what seemed to be meaningful interaction, but without any true understanding or intelligence behind it. One of the most famous scripts Eliza used was the Rogerian psychotherapist model, where the program would respond to a user’s input with open-ended questions and reflections, mimicking the style of a non-directive therapist and creating the illusion of understanding
Dr. Slack was intrigued by ELIZA. He saw ELIZA not just as a technical novelty but as a tool that demonstrated the therapeutic potential of human-computer interaction.
1.5. Views on Soliloquy and Mental Health
Dr. Slack maintained that soliloquy—talking to oneself—whether facilitated by a computer or not, could be a valuable tool for mental health. He challenged the prevailing notion that self-talk was a sign of mental illness. In his own words:
“Contrary to the common notion that soliloquy is a manifestation of mental illness, we believe that it is normal behavior—behavior that serves to help maintain emotional equilibrium.”—Warner Slack, MD
By this, he suggested that engaging in self-directed conversation can help individuals process emotions and thoughts, contributing to psychological well-being.
1.6. Warner Slack’s Legacy and Impact
Dr. Slack’s work was foundational in shifting the healthcare paradigm towards greater patient involvement. His early adoption of computers in medicine paved the way for electronic health records, telemedicine, and patient portals that are commonplace today. His emphasis on patient autonomy and direct interaction with technology anticipated many of the digital health innovations that emerged in the late 20th and early 21st centuries.
2. 1959 and Beyond: Pioneering Machine Learning for Diagnosis—From Nash to Zworykin
Soon after the dawn of the computer age, in the late 1950s, physicians at Cornell Medical School and Mt. Sinai Hospital and engineers at IBM created a mechanical diagnostic aid inspired by the diagnostic slide-rule of British physician F.A. Nash, in which a series of cardboard strips representing symptoms and signs could be lined up to produce a differential diagnosis. (Source: Nash FA. Differential diagnosis: an apparatus to assist the logical faculties. Lancet 1954;263:874–875, p. 874.)
Replacing Nash’s slide rule with a deck of punch cards containing details from clinical history, physical examination, peripheral blood smear, and bone marrow biopsy, the team tried to build a machine that could calculate all hematologic diagnoses. (Source: Greene JA, Lea AS. Digital Futures Past - The Long Arc of Big Data in Medicine. N Engl J Med. 2019 Aug 1;381(5):480-485. doi: 10.1056/NEJMms1817674. PMID: 31365808; PMCID: PMC7928218.)
It was a short step from punch cards to the new IBM digital computer, which the Cornell team saw as a vehicle for more open-ended algorithmic machine learning: information gathered in one analysis could automatically feed back and change the conditions of future analyses. Presenting their work at the first IBM Medical Symposium in 1959, hematologist B.J. Davis claimed that computers could use “raw data,” rather than textbooks, to diagnose disease. (Source: Greene JA, Lea AS. Digital Futures Past - The Long Arc of Big Data in Medicine. N Engl J Med. 2019 Aug 1;381(5):480-485. doi: 10.1056/NEJMms1817674. PMID: 31365808; PMCID: PMC7928218.)
After being fed data of hematologic diagnosis, their IBM 704 computer lumped and split the symptoms and signs into clusters. Reassuringly, the resulting diagnoses largely matched those in textbooks. (Source: Davis BJ. The application of computers to clinical medical data (including machine demonstration). In: Proceedings of the 1st IBM medical symposium. Endicott, NY: IBM, 1959:179–184, p. 185.) Davis and his colleagues saw a future for machine learning in medicine: using a series of feedback loops, machines could organize and modify their own tables of disease on the basis of new data, perhaps more accurately than physicians could. (Source: Lipkin M Correlation of data with a digital computer in the differential diagnosis of hematological diseases. IRE T Med Electron. 1960;ME-7:243–246, p. 246.)
Another Cornell team, led by psychiatrist Keeve Brodman, MD, developed a computerized diagnostic system tested on nearly 6000 outpatients. (Source: Brodman K, van Woerkom AJ. Computer-aided diagnostic screening for 100 common diseases. JAMA 1966;197:901–905.) By 1959, Brodman believed that
“The making of correct diagnostic interpretations of symptoms can be a process in all aspects logical and so completely defined that it can be carried out by a machine.”—Keeve Brodman, MD, a psychiatrist and computer researcher, AMA Archives of Internal Medicine, May 1959.
“What is the character of the error when a diagnosis is made which is not correct? If a patient with flat feet is simply not so diagnosed, this is one type of error, but if the machine reads, ‘respiratory tuberculosis inactive,’ it’s another.”—AMA Archives of Internal Medicine, May 1959.
(Source: Brodman K, van Woerkom AJ, Erdmann AJ Jr, Goldstein LS. Interpretation of symptoms with a data-processing machine. AMA Arch Intern Med 1959;103:116–122, p. 116.)
Keeve Brodman, MD (1906–1979) was a psychiatrist and researcher known for his contributions to medical diagnostics, particularly through the development of the Cornell Medical Index (CMI) in the 1940s. The CMI was a standardized health questionnaire consisting of 195 yes-no questions designed to capture a patient's overall health status. It became a widely used tool in both clinical settings and research, streamlining the collection of patient history and facilitating early attempts at computerized medical diagnostics. (Sources: Lea, Andrew. Computerizing Diagnosis: Keeve Brodman and the Medical Data Screen. Isis 110 (2):228-249, 2019; Max Planck Institute for the History of Science.)
Brodman’s work laid the groundwork for later efforts to integrate computers into the diagnostic process. He was involved in developing the Medical Data Screen (MDS), which aimed to use computerized methods to interpret patient data and aid in diagnostic decision-making. (Sources: Weill Cornell Medicine Library, Max Planck Institute for the History of Science.)
At Georgetown and the University of Rochester, Robert Ledley and Lee Lusted used computational systems to introduce Bayesian logic into clinical diagnosis.
In particular, Lee Lusted, a University of Rochester radiologist and the founder of the Society for Medical Decision Making, predicted the role of computers in radiology in 1959.
“An electronic scanner-computer [will] look at chest photofluorograms, to separate the clearly normal chest films from the abnormal chest films. The abnormal chest films would be marked for later study by the radiologists.”—Lee Lusted, a University of Rochester radiologist and the founder of the Society for Medical Decision Making, 1959
(Source: Lusted LB. Logical analysis in roentgen diagnosis. Radiology. 1960 Feb;74:178-93. doi: 10.1148/74.2.178. PMID: 14419034.)
Other scientists followed suit, working to develop Bayesian systems that could diagnose problems ranging from congenital heart disease to causes of acute abdominal pain. (Sources: Warner HR, Toronto AF, Veasy LG. Experience with Bayes’ theorem for computer diagnosis of congenital heart disease. Ann NY Acad Sci 1964;115:558–567; de Dombal FT, Leaper DJ, Staniland JR, McCann AP, Horrocks JC. Computer-aided diagnosis of acute abdominal pain. BMJ 1972;2: 9–13.)
Researchers aiming to teach diagnosis to these “electronic brains” were divided over the value of replicating diagnostic practices of the “nonelectronic brains” of master physicians versus allowing computers to use distinct logical pathways.
“There is no reason to make the machine act in the same way the human brain does, any more than to construct a car with legs to move from place to place.”—Robert W. Taylor, computer scientist at IBM and director of ARPA’s Information Processing Techniques Office.
(Sources: Ledley RS. Using electronic computers in medical diagnosis. IRE T Med Electron 1960;ME-7:274–280, pp. 279–280; Taylor R, comments on Ledley RS. Using electronic computers in medical diagnosis. IRE T Med Electron 1960;ME-7:274–280, pp. 279–280.)
Yet, the idea of a diagnostic process beyond human understanding was deeply unsettling. Almost 70 years later, it remains a source of discomfort. Today, DARPA—the successor to ARPA—is working hard to demystify AI through its XAI (Explainable Artificial Intelligence) project, aiming to make these complex systems more transparent and understandable.
In 1964, the X-Ray researcher and television inventor Vladimir Zworykin warned that medical data were accumulating at a pace exceeding physicians’ cognitive capacity.
“Not only was the amount of information available in hospital records and medical literature far too large to be encompassed by the memory of any single man, but the conventional techniques of abstracting, summarizing, and indexing cannot provide the physician with the needed knowledge in a form readily accessible in his practice.”—Vladimir Zworykin, an X-Ray researcher and inventor of television, 1964.
Fortunately, the digital computer—which had shrunk from the room-sized ENIAC of the 1940s to the refrigerator-sized IBM mainframe of the 1950s to the “minicomputers” of the 1960s—could deploy algorithmic search techniques at superhuman speeds.
“It is thus quite reasonable to think of electronic memories as effective supplements and extension of the human memory of the physician.”—Vladimir Zworykin, an X-Ray researcher and inventor of television, 1964.
(Source: Zworykin VK. New frontiers in medical electronics: electronic aids for medical diagnosis. Paper delivered at the 39th Anniversary Congress, Pan-American Medical Association, February 19 1964, p. 3. Hagley Museum & Library, Zworykin Vladimir K. Papers (VZP), 1908–1981, David Sarnoff Research Center records (Accession 2464.09), box M&A 78, folder 59.)
Side note:
Vladimir Zworykin’s story is nothing short of fascinating. And, oddly enough, I have my own “six degrees of Kevin Bacon”-style connection to him. Let me explain…
Zworykin is credited as one of the pioneers of television in the 1920s, notably for developing systems that transmitted and received images using cathode ray tubes. However, his journey began much earlier, thanks to his mentor, Professor Boris Rosing, at the St. Petersburg Institute of Technology. It was Rosing who first sparked Zworykin’s interest in television and X-rays. In Zworykin’s personal correspondence, he recounts assisting Rosing with early experiments in the basement of Rosing’s private lab at the School of Artillery in St. Petersburg. Rosing had already filed a patent for a television system in 1907, which featured a rudimentary cathode ray tube as the receiver and a mechanical device as the transmitter. By 1911, he had refined the design and conducted one of the first successful demonstrations of this kind of technology.
Try setting up a “private lab” in modern-day Russia. 🙄 You’d be behind bars by sunrise.
Family Guy nailed its take on Russia. The parallels to reality are uncanny. 👏"
When World War I erupted in 1914, Zworykin enlisted and served in the Russian Signal Corps. He later secured a position with Russian Marconi, where he tested radio equipment produced for the Russian Army. In 1917, the Russian Civil War broke out, pitting the Bolsheviks, led by Vladimir Lenin, against the Whites, commanded by Admiral Alexander Kolchak. Zworykin was hopeful that Kolchak’s forces could defeat the Bolsheviks and liberate Russia from communism. However, in 1918, Zworykin left Russia and emigrated to the United States. He briefly returned to Omsk, the capital of Kolchak’s government, in 1919, traveling via Vladivostok.
Around 1919, Zworykin relocated to Paris, where he worked for a couple of years with Paul Langevin, a renowned physicist often regarded as a founding father of ultrasound and electrodynamics. His time with Langevin expanded Zworykin’s understanding of X-rays and other advanced scientific techniques. Langevin’s mentorship played a crucial role in shaping Zworykin’s grasp of physics and research methodologies, which would later influence his groundbreaking work in television technology and electronics.
Paul Langevin developed the world famous Langevin Equation, the fundamental equation of non-equilibrium systems. Langevin equation describes the motion of particles under the influence of friction and random forces, laying the groundwork for Brownian motion. The modern era in the theory of Brownian motion began with Albert Einstein. Langevin’s work in paramagnetism and diamagnetism also influenced the understanding of how materials respond to magnetic fields.
Fun fact: Named after a twice Nobel Prize winner Polish-French physicist Marie Curie, the Curie Law is actually a special case of the Langevin Equation.
In the early 1920s, Vladimir Zworykin moved permanently to the United States and eventually became the head of television development for the Radio Corporation of America (RCA).
My connection to Zworykin comes through the Langevin Equation. 😊 One important field of study that evolved from the Langevin Equation is Stochastic Thermodynamics. This emerging field in statistical mechanics leverages stochastic variables to better understand the non-equilibrium dynamics found in various microscopic systems, such as colloidal particles, biopolymers (e.g., DNA, RNA, and proteins), enzymes, and molecular motors.
Speaking of Stochastic Thermodynamics, the Langevin Equation (also known as the Langevin Law) is at the core of my recent publication, which I co-authored with two colleagues—one of whom is my father. We developed a model that explores the interaction between a uniform magnetic field and a cylindrical ferrofluid layer. The study has significant implications for the healthcare industry by aiming to reduce the risk of stroke triggered by magnetic (solar) storms. Essentially, it proposes a lightweight helmet with a thin ferrofluid layer covering the entire surface. This device could improve the quality of life for the approximately 6% of people worldwide who suffer from debilitating headaches caused by magnetic (solar) storms, allowing them to live a normal life.
Anyway, I digress…
Attempts to automate diagnostic reasoning also required implicit valuations of varieties of medical work—separating monotonous, mechanizable processes from those that were complex and inextricably human. Advocates argued that medical computing could ease the growing physician shortage and allow doctors to focus more on human interaction. (Source: Bane F Physicians for a growing America. Washington, DC: US Government Printing Office, 1959.)
Critics objected, as Zworykin put it, “to the misconception that we were trying to replace the doctor by cold, hard, calculating machine,” fearing disruption of personal bonds between doctors and patients. (Sources: Zworykin VK. Welcome—introduction to the conference. IRE T Med Electron 1960;ME-7:239; Brodman K to Wander A, September 27, 1967. Medical Center Archives of New York Presbyterian/Weill Cornell, The Keeve Brodman, MD (1906–1979) Papers, box 4, folder 2; “Computers programmed to sort routine symptoms,” n.d. Medical Center Archives of New York Presbyterian/Weill Cornell, The Ralph Engle, MD (1920–2000) Papers, box 8, folder 2.)
“The misconception [was] that we were trying to replace the doctor by cold, hard, calculating machine.”—Vladimir Zworykin, an X-Ray researcher and inventor of television, at the The Rockefeller Institute for Medical Research’s conference on Diagnostic Data Processing, New York, 1960.
Even as some physicians hailed the computer’s rote nature as a solution to errors arising from being “too human,” others warned that the computer would become an additional source of error, whether owing to bugs in its code or to human biases underlying that code.
(Source: McTernan E, Crocker D. Push-button medicine is no pipe dream! Hospital Physician January 1969; 85.)
As Medical World News explained in 1967,
“Diagnosis requires judgment, and a computer must rely on its human programmers to supply that judgment. The ‘electronic brain’ is no brain at all. It has been called ‘an idiot machine,’ capable of generating errors at the same fabulous rate at which it generates correct answers.”—Medical World News, July 14, 1967.
(Source: Medicine faces the computer revolution: electronic ‘brains’ are heralding a new epoch of improved diagnosis, timelier treatment, and far less medical paper work. Medical World News, July 14, 1967; 46–55, on p. 47.)
Today, there are similar fears that machine learning might—instead of eliminating lapses in human judgment—harden errors and biases into rhythms of care. (Source: Mullainathan S, Obermeyer Z, Does machine learning automate moral hazard and error? Am Econ Rev 2017;107:476–480.)
Perhaps the most sobering appraisals of early computerized diagnostic systems came from engineers. By 1969, the Cornell team conceded that the most machine learning could promise was an algorithm for hematologic diagnoses—a relatively easy diagnostic area to conceive as a logical tree—that matched existing textbooks. Ironically, even some of the greatest successes of 1960s diagnostic computing—such as the automated readings produced by electrocardiograms—have become so naturalized that they’re no longer thought of as computer diagnosis. (Source: Progress Report: Electronic Data Processing in Hematology, Grant No. AM-06857–03, U.S. Public Health Services, National Institutes of Health, January 1963, p. 2. Medical Center Archives of New York Presbyterian/Weill Cornell, The Ralph Engle, MD (1920–2000) Papers, box 14, folder 11.)
Pioneers of computerized diagnosis ran up against inherent complexities. Zworykin’s team, for example, grew exasperated by uncertainty and medical heterogeneity. In seeking raw data for their system by reviewing the medical literature, they encountered a messy world of information in which researchers and clinicians reported data according to their own preferences, particularities, and biases. Perhaps more concerning,
If medical experts themselves often disagreed about correct diagnoses, how could one know when the computer was right? Developers found it difficult to define gold standards against which to evaluate their systems’ accuracy.
(Source: Buchanan BG, Shortliffe EH. The problem of evaluation. In: Buchanan BG, Shortliffe EH, eds. Rule-based expert systems: the MYCIN experiments of the Stanford Heuristic Programming Project. Reading, MA: Addison-Wesley, 1984, pp. 571–588.)
Such challenges persist today, both in designing adequate evaluation studies and in generating diagnostic “ground truth” for machine-learning algorithms. (Source: Schenthal JE, Sweeney JW, Nettleton WJ, Yoder RD. Clinical applications of electronic data processing apparatus III: system for processing medical records. JAMA 1963;186:101–105, p. 101.)
Some failings of early computer diagnosis pertained to technological and informational limitations of the time, but, more than half a century later, many fundamental challenges identified back then have yet to be overcome. (Source: Greene JA, Lea AS. Digital Futures Past - The Long Arc of Big Data in Medicine. N Engl J Med. 2019 Aug 1;381(5):480-485. doi: 10.1056/NEJMms1817674. PMID: 31365808; PMCID: PMC7928218.)
Also in the late 1950s, early biomedical researchers began exploring the possibility of using computers to investigate and solve problems in biology and medicine. Some of those studies were ultimately directed toward the development of systems for computer-based medical diagnosis (Sources: Ledley, 1959; Vandenberg, 1960; Weinrauch and Hetherington, 1959).
These early diagnostic systems, often referred to as “expert systems in medicine,” used patients’ symptoms and clinical data to emulate the diagnostic reasoning of human physicians.
An expert system is an AI program designed to (a) provide expert-level solutions to complex problems, (b) be understandable, and (c) be flexible enough to easily incorporate new knowledge. (Source: Buchanan BG, Shortliffe EH. The problem of evaluation. In: Buchanan BG, Shortliffe EH, eds. Rule-based expert systems: the MYCIN experiments of the Stanford Heuristic Programming Project (Chapter 1). Reading, MA: Addison-Wesley, 1984, page 3.)
Medical Artificial Intelligence is primarily concerned with the construction of AI programs that perform diagnosis and make therapy recommendations. Unlike medical applications based on other programming methods, such as purely statistical and probabilistic methods, medical AI programs are based on symbolic models of disease entities and their relationship to patient factors and clinical manifestations. (Clancey, W. J., & Shortliffe, E. H. (1984). Readings in Medical Artificial Intelligence: The First Decade. Addison-Wesley.)
The first computer-aided diagnosis (CAD) approaches were based on these expert systems in the 1950s. This was followed by CAD approaches based on intelligent systems in the 1960s and 1970s.
3. 1960s: PROMIS, the First Touch-Screen Computer-Based Medical Record System
Judy Faulkner wants you to think Epic invented the electronic medical record, a narrative she’s pushed with her famous meme. But nothing could be farther from the truth. There was a whole universe of innovation long before Epic showed up.
The Problem-Oriented Medical Information System (PROMIS) was one of the earliest computer-based medical record systems, developed to improve patient care through enhanced data management. Emerging as a computer-based patient-oriented medical record (POMR), a prototype of PROMIS was created in 1967 by Jan Schultz and Lawrence L. Weed, M.D., initially at the Cleveland Metropolitan General Hospital (CMGH). In 1969, the project moved to the University of Vermont’s Medical Center Hospital of Vermont (MCHV) in Burlington. PROMIS aimed to transform how medical information was recorded, stored, and utilized by healthcare professionals. By implementing a structured, problem-oriented approach, it laid the groundwork for modern electronic health records (EHRs) and made a lasting impact on the field of medical informatics.
The system utilized touch-screen terminals and a scalable network as early as 1967, which were groundbreaking at the time. PROMIS allowed healthcare providers to input data directly, supporting a problem-oriented medical record (POMR) format. It included features such as dynamic flow sheets and the ability to retrieve information by problem, source, or time. The development involved collaboration with various experts, and the lab where it was created, located at the University of Vermont, became known as the “Promised Land.”
How PROMIS Worked
Purpose
Structured Medical Records: PROMIS was designed to systematically organize patient medical records using the Problem-Oriented Medical Record (POMR) framework, allowing for easier tracking of patient information and care plans.
Clinical Decision Support: It aimed to assist healthcare providers in diagnosing and treating patients by providing organized and accessible medical information. This was achieved through the integration of medical guidelines and protocols within the system.
Improved Communication: PROMIS sought to enhance communication among healthcare teams by standardizing how patient information was recorded and shared, ensuring continuity of care.
Knowledge Base
Problem Lists: Central to PROMIS was the creation of comprehensive problem lists for each patient, documenting all identified medical issues. This allowed for better tracking and management of multiple conditions.
Standardized Data Entry: PROMIS encouraged consistent recording of patient data, including medical history, symptoms, diagnoses, treatments, and outcomes. This structured approach minimized the risk of errors and omissions.
Integrated Medical Knowledge: The system included medical guidelines and protocols to support clinical decision-making. For instance, it provided recommendations for diagnostic and therapeutic actions based on the recorded data.
User Interaction
Input Mechanism: Clinicians entered patient data using innovative touch-screen terminals, which were considered advanced for that era. Initially, Jan Schultz and Larry Weed experimented with an IBM 1440 computer, the industry standard at the time, but they found the process to be cumbersome and slow. Instead, they adopted the Digiscribe, a touch-screen CRT terminal developed by Control Data Corporation, which allowed for more efficient data entry.
User-Friendly Interface: PROMIS featured menu-driven interfaces and prompts to guide users through data entry, making it easier to use and reducing the likelihood of input errors. The touch-screen technology was ahead of its time, setting a standard for future healthcare systems.
Output: The system generated organized patient records, updated problem lists, and treatment plans, all of which were easily accessible to healthcare teams, thus facilitating coordinated care.
Technological Foundations
Programming Language
FORTRAN: Before 1973, the existing system couldn’t handle the vast amount of data, prompting the development of a new architecture. A network of minicomputers with interconnected nodes was designed, and new hardware and software were procured. FORTRAN was installed on the V77-400 Varian Data Machines minicomputers within this network. The newly developed PROMIS system proved to be more efficient and capable of managing larger data volumes.
Knowledge Representation
Problem-Oriented Structure: PROMIS employed the POMR approach to organize medical data around specific patient problems. This was revolutionary in shifting focus from data entry to problem-solving.
Modular Design: The modular design allowed for flexibility in adding or updating patient information without overhauling the entire system. This adaptability was key to managing complex medical cases.
Computational Environment
Mainframe Computers: PROMIS operated on mainframe systems capable of handling substantial amounts of data processing and storage. It was designed to be scalable, allowing multiple terminals to be connected across a hospital.
Touch-Screen Terminals: Utilizing custom-built touch-screen devices, PROMIS enabled direct interaction, reducing the need for intermediary data entry personnel. These terminals were responsive, processing 70% of selections within 250 milliseconds, making the system efficient for real-time use.
Operational Workflow
Data Entry: Healthcare providers input patient information, including medical history and current complaints, via touch-screen terminals. This direct data entry ensured accuracy and timeliness.
Problem List Creation: PROMIS organized this data into a structured problem list, highlighting each medical issue. This list was continuously updated, reflecting the patient’s ongoing condition and treatments.
Care Planning: For each problem, clinicians could develop diagnostic and therapeutic plans within the system. The system’s integration of clinical guidelines made this process more straightforward and consistent.
Progress Monitoring: PROMIS enabled continuous updates to the patient’s status, treatments, and outcomes, ensuring that real-time data was available to all healthcare providers involved in the patient’s care.
Information Retrieval: The system allowed easy access to patient records, facilitating informed decision-making and coordinated care. It could present data in various ways, such as problem-oriented, time-oriented, or source-oriented, enhancing its utility.
Technological Innovations
Early Adoption of Touch-Screen Technology: PROMIS was among the first systems to implement touch-screen interfaces in healthcare. Despite being radical at the time, this innovation improved usability and set a precedent for future systems.
Implementation of POMR in Computing: Successfully translating the POMR method into a computerized format was a significant achievement, making patient data more accessible and manageable.
Interactive User Interface: The system’s design prioritized ease of use, encouraging adoption by medical staff. Its menu-driven, responsive interface was a leap forward in healthcare technology.
Integration of Clinical Guidelines: By incorporating medical knowledge bases, PROMIS assisted with diagnosis and treatment planning, providing clinicians with actionable insights.
Impact and Legacy
Pioneering Electronic Medical Records: PROMIS significantly influenced the development of EHRs by demonstrating the benefits of computerized medical records. Its focus on problem-oriented care became a guiding principle for future systems.
Enhanced Patient Care: By organizing patient information comprehensively and making it readily accessible, PROMIS improved the quality of care. Physicians could quickly retrieve relevant data, enabling more accurate diagnoses and timely interventions.
Educational Influence: PROMIS served as a model for teaching the importance of structured medical records in medical education. It illustrated how systematic documentation could improve patient outcomes and streamline care processes.
Advancement of Medical Informatics: PROMIS contributed to the evolution of medical informatics, inspiring future innovations and collaborations between clinicians and technologists.
Recognition of Challenges: The development and implementation of PROMIS highlighted issues such as the need for user training, system costs, and technical limitations, which guided improvements in subsequent healthcare IT systems.
Operational Example
Consider a physician treating a patient with diabetes and hypertension. Using PROMIS, the physician enters the patient’s medical history, current symptoms, and laboratory results into the system via a touch-screen terminal. PROMIS organizes this information into a problem list, identifying diabetes and hypertension as primary issues. For each problem, the physician documents diagnostic plans (e.g., ordering HbA1c tests) and therapeutic interventions (e.g., adjusting medications). The system allows for continuous updates, so as the patient’s condition changes, the physician can modify the problem list and treatment plans accordingly, ensuring coordinated and effective care.
Technological Advancements Inspired by PROMIS
Development of EHR Systems: PROMIS’s success laid the groundwork for the widespread adoption of EHRs in healthcare settings worldwide.
User Interface Design: The user-friendly interfaces in PROMIS influenced the creation of more intuitive and efficient interfaces in later medical software applications.
Standardization in Healthcare Documentation: PROMIS promoted the adoption of standardized medical record-keeping practices, which became a norm in the healthcare industry.
Interdisciplinary Collaboration: The development of PROMIS encouraged collaboration between clinicians and technologists, fostering innovations that bridged healthcare and information technology.
PROMIS stands as a monumental step in the digitization of healthcare records. Its innovative use of technology and commitment to improving patient care through better information management have left a lasting legacy in the medical community. By demonstrating the practical benefits of computerized medical records, PROMIS paved the way for the advanced EHR systems that are integral to modern healthcare delivery.
References for Further Reading
Weed, L. L. (1968). Medical Records, Medical Education, and Patient Care: The Problem-Oriented Record as a Basic Tool. Case Western Reserve University Press.
Schultz, J. (1986, January 10). A history of the PROMIS technology—An effective human interface [Conference presentation]. ACM Conference on the History of Personal Workstations, Xerox Palo Alto Research Center (PARC), Palo Alto, CA. YouTube.
Note: Everyone, especially those working for an EHR company, should watch this 1986 presentation by Jan Schultz, one of the developers of PROMIS, in Palo Alto. During the presentation, Jan showcases two demos of the PROMIS system from September 1974, one lasting 13 minutes and the other 11 minutes. It’s humbling to see that they had already implemented touch-screen technology back then. Given that the computing power of the entire system was just a fraction of what we have in today’s smartphones, the innovations developed at the PROMIS project were truly remarkable.
Collen, M. F. (1995). A history of medical informatics in the United States, 1950 to 1990. American Medical Informatics Association.
McDonald, C. J. (1997). The evolution of Intel-based hospital information systems at the Regenstrief Institute. MD Computing, 14(2), 112–118.
Schultz, J., & Rees, J. (2019). The “PROMIS” of Computer-Based Medical Records. National Library of Medicine Archives.
4. 1960s: CASNET, the First AI System for Medical Diagnosis and Treatment
CASNET (Causal Associational Network) was one of the earliest expert systems developed for medical diagnosis. Created in the late 1960s and early 1970s by Dr. Casimir A. Kulikowski and Dr. Sholom M. Weiss at Rutgers University, CASNET was designed to assist in the diagnosis and management of glaucoma—a group of eye conditions that can lead to blindness. Although developed before the 1980s, CASNET’s foundational principles and technological innovations significantly influenced the development of later expert systems during the 1980s and beyond.
How CASNET Worked
Purpose
Diagnostic Assistance: Aimed to aid ophthalmologists in diagnosing glaucoma by analyzing patient symptoms, medical history, and examination results.
Treatment Recommendations: Provided suggestions for appropriate treatment plans based on the diagnosed condition and patient-specific factors.
Knowledge Base
Causal Network Framework: Contained a comprehensive causal network that encapsulated expert knowledge about glaucoma, including symptom presentation, disease progression, and treatment protocols.
Data Representation: Utilized causal associations to correlate patient data with possible diagnoses and treatment options.
Inference Engine
Combined Reasoning: Employed both forward and backward chaining, starting with available patient information and using the causal network to infer diagnoses and treatment plans.
Probabilistic Reasoning: Incorporated probabilities to handle uncertainties inherent in medical diagnoses, enhancing the accuracy of its recommendations.
User Interaction
Input Mechanism: Physicians entered patient data through an interface, inputting symptoms, examination findings, and medical history.
Output: Provided diagnostic suggestions and recommended treatments, along with explanations of the reasoning behind each recommendation.
Technological Foundations
Programming Language
FORTRAN and SNOBOL: The consultation program was interactive, running in 35K words of memory on a DEC 10 or 20 computer under either the TOPS-20 or TENEX operating systems. Because of speed and efficiency considerations, it is written in FORTRAN. Modifications and updating of the glaucoma model are carried out by interaction with the separate model-building program, written in SNOBOL. This program checks the model for consistency and compiles it so that it will run efficiently under CASNET/Glaucoma.
Knowledge Representation
Causal Networks: Central to CASNET’s functionality, allowing it to model the relationships between different medical concepts through causal links.
Probabilistic Models: Used to represent the likelihood of certain diagnoses given specific symptoms and findings, facilitating decision-making under uncertainty.
Computational Environment
Mainframe Computers: Ran on mainframe systems, which provided the necessary computational power for processing complex networks and large amounts of patient data.
Operational Workflow
Data Entry: The physician inputs patient-specific data, including symptoms, medical history, and examination findings such as intraocular pressure measurements.
Causal Reasoning: CASNET analyzes the input data within its causal network to evaluate the likelihood of various disease states associated with glaucoma.
Diagnosis Generation: Based on the analysis, CASNET infers potential diagnoses, categorizing them by type (e.g., open-angle glaucoma, angle-closure glaucoma) and severity.
Treatment Recommendations: Suggests appropriate treatment plans tailored to the diagnosed type of glaucoma, considering factors like patient age, disease progression, and risk factors.
Feedback Mechanism: Physicians could provide feedback on the system’s recommendations, allowing for manual adjustments and iterative refinement of diagnostic suggestions.
Technological Innovations
Early Use of Causal Networks: CASNET pioneered the use of causal networks for medical diagnosis, setting the stage for future expert systems that utilize complex knowledge representations.
Automated Diagnostic Reasoning: Demonstrated the feasibility of automating diagnostic reasoning processes, reducing cognitive load on physicians and minimizing diagnostic errors.
Integration of Treatment Protocols: Not only focused on diagnosis but also integrated treatment recommendations, providing a holistic decision support tool.
Handling Uncertainty: Incorporated probabilistic reasoning to manage uncertainties in medical diagnosis, enhancing the reliability of its outputs.
Impact and Legacy
Foundational Influence: CASNET’s development laid the groundwork for subsequent medical expert systems, influencing the design and functionality of later systems in healthcare.
Academic Contributions: Served as a key case study in the early exploration of AI applications in medicine, contributing significantly to academic discourse and research in both artificial intelligence and medical informatics.
Demonstration of AI Potential: Validated the potential of artificial intelligence to augment medical diagnostics, encouraging further investment and research into AI-driven healthcare solutions.
Limitations and Challenges: Highlighted early challenges in expert system development, such as knowledge acquisition, maintenance of complex networks, and user interface design, informing future improvements in computer-aided diagnosis systems.
Evolution into Advanced Systems: Inspired the incorporation of more sophisticated reasoning techniques, enhanced probabilistic models, and user-friendly interfaces in later expert systems during the 1980s and beyond.
Operational Example
Imagine an ophthalmologist using CASNET to diagnose a patient presenting with symptoms like increased intraocular pressure, visual field loss, and optic nerve damage. The physician inputs these symptoms and examination findings into CASNET, which then analyzes the data within its causal network to identify patterns indicative of different types of glaucoma. CASNET might suggest a diagnosis of open-angle glaucoma and recommend a specific treatment plan based on the patient’s history and the severity of the condition, providing the physician with a reasoned basis for the recommendation.
Technological Advancements Inspired by CASNET
Knowledge Engineering: Emphasized the importance of structured knowledge acquisition and representation, leading to more advanced knowledge engineering techniques in later systems.
Human-Computer Interaction: Highlighted the need for more intuitive and user-friendly interfaces, influencing the design of graphical user interfaces in subsequent computer-aided diagnosis systems.
Interdisciplinary Collaboration: Fostered collaboration between computer scientists and medical professionals, promoting interdisciplinary approaches to developing effective expert systems.
CASNET stands as a pioneering effort in the application of artificial intelligence to medical diagnosis. Its development in the late 1960s and early 1970s showcased the potential of expert systems to assist healthcare professionals, influencing the trajectory of computer-aided diagnosis system development throughout the 1980s and beyond. By addressing both diagnostic and treatment recommendations, CASNET provided a comprehensive decision support tool, paving the way for more sophisticated and specialized expert systems that continue to evolve in today’s healthcare landscape.
References for Further Reading
Kulikowski, C. A., & Weiss, S. M. (1982). Representation of expert knowledge for consultation: the CASNET and EXPERT projects. In P. Szolovits (Ed.), Artificial Intelligence in Medicine (pp. 21-55). Westview Press.
Weiss, S. M., & Kulikowski, C. A. (1984). A Practical Guide to Designing Expert Systems. Rowman & Allanheld.
Clancey, W. J., & Shortliffe, E. H. (1984). Readings in Medical Artificial Intelligence: The First Decade. Addison-Wesley.
5. 1970s and 1980s: The Rise of Artificial Intelligence in Medicine through Expert Systems
The 1970s and 1980s were a formative period for the application of expert systems in medical diagnosis. Systems like INTERNIST-I, CADUCEUS, QMR, CADIAG, MYCIN, and DXplain showcased the potential of artificial intelligence to augment medical expertise, improve diagnostic accuracy, and streamline patient care. While many of these systems were constrained by the technological limitations of their time, their development provided critical insights and laid the groundwork for the sophisticated AI-driven healthcare solutions we benefit from today.
(Sources: Rule-based expert systems : the MYCIN experiments of the Stanford Heuristic Programming Project / edited by Bruce G. Buchanan, Edward H. Shortliffe. Reading, Mass.: Addison-Wesley, c1984; Szolovits, P. Artificial Intelligence in Medicine. Westview Press, Boulder, Colorado, 1982; Juri Yanase, Evangelos Triantaphyllou, A systematic survey of computer-aided diagnosis in medicine: Past and present developments, Expert Systems with Applications, Volume 138, 2019, 112821, ISSN 0957-4174, https://doi.org/10.1016/j.eswa.2019.112821; Hirani R, Noruzi K, Khuram H, Hussaini AS, Aifuwa EI, Ely KE, Lewis JM, Gabr AE, Smiley A, Tiwari RK, Etienne M. Artificial Intelligence and Healthcare: A Journey through History, Present Innovations, and Future Possibilities. Life (Basel). 2024 Apr 26;14(5):557. doi: 10.3390/life14050557. PMID: 38792579; PMCID: PMC11122160.)
5.1. INTERNIST-I: The First AI Model for Internal Medicine Diagnostic Reasoning
INTERNIST-I was a pioneering medical expert system developed in 1972-73 at the University of Pittsburgh by Dr. Jack D. Myers, Dr. Harry E. Pople Jr., and Dr. Randolph A. Miller. Designed to assist physicians in diagnosing complex cases in internal medicine, INTERNIST-I was one of the earliest attempts to model the diagnostic reasoning of expert clinicians using artificial intelligence. Its development significantly influenced the design of subsequent medical expert systems in the 1980s and beyond.
How INTERNIST-I Worked
Purpose
Diagnostic Assistance: Aimed to support physicians in diagnosing complex internal medicine cases by analyzing patient symptoms, history, and laboratory data.
Educational Tool: Served as a teaching aid for medical students and residents, illustrating the diagnostic reasoning process.
Knowledge Base
Comprehensive Disease Profiles: Contained detailed information on over 500 diseases in internal medicine, including their associated symptoms, signs, and laboratory findings.
Manifestation-Disease Relationships: Represented the relationships between diseases and their manifestations using a network of weighted connections, indicating the strength and nature of each association.
Inference Engine
Heuristic Reasoning: Employed heuristic algorithms to simulate the problem-solving strategies of expert clinicians.
Scoring System: Used a scoring mechanism to evaluate and rank potential diagnoses based on the presence and importance of clinical findings.
User Interaction
Input Mechanism: Physicians entered patient data through a textual interface, including symptoms, physical examination findings, and laboratory results.
Output: Provided a ranked list of possible diagnoses along with explanations of how each diagnosis correlated with the patient’s findings.
Technological Foundations
Programming Language
LISP: INTERNIST-I was implemented in LISP, a language favored in AI research for its proficiency in handling symbolic reasoning and complex data structures.
Knowledge Representation
Entity-Relationship Model: Organized medical knowledge using entities (diseases) and their relationships with manifestations (symptoms and signs).
Weighted Associations: Each manifestation-disease link had an assigned weight reflecting the likelihood or significance of the association.
Computational Environment
Mainframe Computers: Operated on mainframe systems of the era, which provided the necessary computational power for processing extensive medical data and complex reasoning algorithms.
Operational Workflow
Data Entry: The physician inputs detailed patient information, including symptoms, signs, and laboratory results.
Hypothesis Generation: INTERNIST-I analyzes the input data to generate a list of potential diseases that could explain the patient’s manifestations.
Scoring and Ranking: Assigns scores to each potential diagnosis based on the weighted associations, ranking them according to their likelihood.
Diagnostic Suggestions: Presents the physician with a ranked list of probable diagnoses, along with explanations and relevant supporting information.
Iterative Refinement: Allows for additional data input and refinement of the diagnostic list as more information becomes available.
Technological Innovations
Simulation of Expert Reasoning: Modeled the diagnostic reasoning process of experienced clinicians using heuristic methods.
Extensive Knowledge Base: One of the first systems to incorporate a vast and detailed medical knowledge base covering a wide range of internal medicine.
Weighted Diagnostic Associations: Introduced the concept of weighting the significance of clinical findings in relation to diseases, improving diagnostic accuracy.
Educational Impact: Enhanced medical education by providing insights into the diagnostic process and reasoning strategies.
Impact and Legacy
Foundation for Future Systems: INTERNIST-I served as the foundation for CADUCEUS and later Quick Medical Reference (QMR), which further refined its knowledge base and reasoning processes.
Advancement of Medical AI: Demonstrated the potential of AI in medicine, encouraging further research and development in diagnostic support systems.
Highlighting Challenges: Exposed limitations such as the difficulty in maintaining and updating extensive knowledge bases, influencing future approaches to knowledge representation.
Interdisciplinary Collaboration: Strengthened collaboration between medicine and computer science, fostering interdisciplinary innovation.
Operational Example
Imagine a patient presenting with fatigue, joint pain, skin rash, and positive antinuclear antibodies. A physician inputs these findings into INTERNIST-I. The system processes the data, matching the patient’s manifestations with its disease profiles. It might generate a ranked list of possible diagnoses such as systemic lupus erythematosus, rheumatoid arthritis, and polymyositis, providing explanations for each based on the weighted associations. The physician can then consider these suggestions, order further tests, and refine the diagnosis.
Technological Advancements Inspired by INTERNIST-I
Development of QMR: INTERNIST-I evolved into the Quick Medical Reference system, which offered enhanced diagnostic reasoning and a more user-friendly interface.
Improved Knowledge Representation: Influenced the adoption of more sophisticated methods for representing medical knowledge, including probabilistic models.
Integration of Probabilistic Reasoning: Inspired the incorporation of Bayesian reasoning and other statistical methods to handle diagnostic uncertainty.
Enhanced User Interfaces: Highlighted the importance of intuitive interfaces, leading to the development of graphical user interfaces in later systems.
INTERNIST-I stands as a landmark achievement in the application of artificial intelligence to medicine. Its innovative approach to simulating expert diagnostic reasoning provided valuable insights and tools for both practicing physicians and medical educators. By addressing the complexities of internal medicine diagnosis, INTERNIST-I paved the way for more advanced and user-friendly expert systems that continue to evolve in the healthcare industry.
References for Further Reading
Pople, H. E., Myers, J. D., & Miller, R. A. (1975). DIALOG: A Model of Diagnostic Logic for Internal Medicine. Proceedings of the 4th International Joint Conference on Artificial Intelligence, 848–855.
Pople, H. E. (1982). Heuristic methods for imposing structure on ill-structured problems: The structuring of medical diagnostics. In P. Szolovits (Ed.), Artificial Intelligence in Medicine (pp. 119-190). Massachusetts Institute of Technology. https://people.csail.mit.edu/psz/ftp/AIM82/ch5.html
Miller, R. A., Pople, H. E., & Myers, J. D. (1982). INTERNIST-1, an experimental computer-based diagnostic consultant for general internal medicine. The New England Journal of Medicine, 307(8), 468–476.
Miller, R. A. (1994). Medical diagnostic decision support systems—past, present, and future: a threaded bibliography and brief commentary. Journal of the American Medical Informatics Association, 1(1), 8–27.
5.2. MYCIN: Bridging Artificial Intelligence and Medical Diagnosis
MYCIN was one of the earliest and most influential expert systems in artificial intelligence, developed in the early 1970s at Stanford University by Dr. Edward H. Shortliffe under the guidance of Dr. Bruce G. Buchanan and others. Designed to assist physicians in diagnosing and treating bacterial infections, particularly bacteremia and meningitis, MYCIN represented a significant advancement in medical AI. Although it was never implemented in clinical practice, MYCIN’s innovative approach to knowledge representation and reasoning under uncertainty had a profound impact on the development of later expert systems during the 1980s and beyond.
The name “MYCIN” was chosen to reflect the system’s focus on medical knowledge related to antibiotics, many of which have names that end in “-mycin” (such as erythromycin and streptomycin).
MYCIN was part of the Stanford Heuristic Programming Project (HPP), along with other projects like DENDRAL, CONGEN, Meta-DENDRAL, and SU/X (later renamed HASP/SIAP). Within the MYCIN project, researchers worked on several nearly distinct subprojects, as shown in the diagram above: Question Answering (QA), Inference (which included certainty factors, or CFs, and the therapy recommendation code), Explanation, Evaluation, and Knowledge Acquisition.
It all began with DENDRAL (short for “Dendritic Algorithm”). DENDRAL, an early expert system, was developed starting in 1965 by artificial intelligence (AI) researcher Edward Feigenbaum and geneticist Joshua Lederberg, both of Stanford University. Heuristic DENDRAL (later simply called DENDRAL) was designed as a chemical-analysis expert system. For example, if the substance being analyzed was a complex compound of carbon, hydrogen, and nitrogen, DENDRAL would hypothesize its molecular structure based on spectrographic data. Its performance was on par with chemists skilled in this task, and the program found applications in both industry and academia.
How MYCIN Worked
Purpose
Diagnostic Assistance: Aimed to help physicians identify the specific bacteria causing an infection by analyzing patient symptoms, medical history, and laboratory results.
Treatment Recommendations: Provided tailored antibiotic therapy suggestions, including dosages and potential drug interactions, based on the diagnosed organism and patient-specific factors.
Knowledge Base
Rule-Based System: Contained over 600 heuristic rules derived from expert knowledge in infectious diseases and antimicrobial therapy.
Medical Expertise: Encoded knowledge about bacteria, antibiotics, infection sites, and patient conditions to facilitate accurate diagnosis and treatment.
Inference Engine
Backward Chaining: Employed backward chaining to start with potential hypotheses (e.g., possible bacteria) and worked backwards to gather supporting evidence from patient data.
Certainty Factors: Introduced the use of certainty factors to handle uncertainty, allowing the system to express degrees of confidence in its conclusions.
User Interaction
Interactive Dialogue: Engaged physicians in a question-and-answer session, prompting for specific patient information as needed.
Explanations: Provided detailed explanations of its reasoning process, including the rules applied and certainty levels, enhancing transparency and trust.
Technological Foundations
Programming Language
LISP Implementation: MYCIN was developed using LISP, a language favored in AI research for its strengths in symbolic processing and manipulation of complex data structures.
Knowledge Representation
Production Rules: Used if-then production rules to represent medical knowledge, allowing for modularity and ease of updating the knowledge base.
Certainty Factors: Pioneered a formalism for representing uncertainty in expert systems, assigning numerical values to express the confidence in both rules and conclusions.
Computational Environment
Time-Sharing Systems: Operated on mainframe time-sharing systems, providing the computational resources necessary for real-time interaction and complex reasoning tasks.
Operational Workflow
Data Entry: The physician inputs patient-specific information, including symptoms, medical history, and laboratory test results such as blood cultures.
Hypothesis Formation: MYCIN generates hypotheses about potential causative bacteria based on the input data.
Interactive Questioning: The system asks additional, targeted questions to gather more detailed information, refining its hypotheses.
Inference Processing: Applies relevant production rules using backward chaining, evaluating certainty factors to handle conflicting or incomplete data.
Diagnosis: Identifies the most probable bacterial organisms responsible for the infection, along with associated certainty levels.
Treatment Recommendations: Suggests appropriate antibiotic regimens, considering factors like bacterial susceptibility, patient allergies, and drug interactions.
Explanation and Justification: Offers explanations for its diagnoses and recommendations, detailing the reasoning steps and evidence considered.
Technological Innovations
Uncertainty Management: Introduced certainty factors, allowing the system to reason effectively under uncertainty, a common challenge in medical diagnosis.
Rule-Based Expert Systems: Established a framework for building expert systems using production rules, influencing numerous subsequent AI applications.
Explanation Facility: Provided users with understandable justifications for its conclusions, setting a precedent for transparency in AI systems.
Knowledge Engineering: Advanced methodologies for eliciting and encoding expert knowledge, contributing to the field of knowledge engineering.
Impact and Legacy
Foundational Influence: MYCIN’s architecture and reasoning methods became a blueprint for future expert systems in medicine and other domains.
Academic Contributions: Generated extensive research and literature, shaping the fields of medical informatics and artificial intelligence.
Expert System Shells: Led to the development of EMYCIN (Empty MYCIN), a general-purpose expert system shell that allowed others to build similar systems in different domains.
Ethical and Practical Discussions: Sparked debates on the ethical implications of AI in healthcare, user acceptance, and the integration of such systems into clinical practice.
Limitations Highlighted: Brought attention to challenges like knowledge base maintenance, user interface design, and the complexities of replicating human expertise.
Operational Example
Imagine a physician faced with a patient exhibiting symptoms of severe infection, such as high fever and low blood pressure. The physician inputs initial findings into MYCIN, which then asks specific questions about symptoms, lab results (e.g., white blood cell count, culture results), and patient history (e.g., allergies, current medications). Using its rule-based system and certainty factors, MYCIN evaluates possible bacterial causes, such as Staphylococcus aureus or Escherichia coli. It then recommends an antibiotic treatment plan, specifying drug choices, dosages, and administration routes, while providing explanations for each recommendation based on the evidence and rules applied.
Technological Advancements Inspired by MYCIN
Expert System Development: Encouraged the creation of other expert systems across various fields, utilizing rule-based reasoning and uncertainty management.
Uncertainty Representation: Influenced the adoption of probabilistic reasoning and fuzzy logic in AI to better handle incomplete or ambiguous data.
Human-Computer Interaction: Emphasized the importance of user-friendly interfaces and explanation facilities, impacting the design of subsequent AI systems.
Knowledge Acquisition Techniques: Advanced strategies for extracting and formalizing expert knowledge, improving the efficiency and effectiveness of building expert systems.
MYCIN stands as a pioneering effort in applying artificial intelligence to complex medical problems. Its development in the 1970s showcased the potential of expert systems to augment physician decision-making, particularly in areas requiring specialized knowledge like infectious diseases. While MYCIN itself was not deployed in clinical settings, largely due to ethical, practical, and technical constraints of the time, its innovations laid critical groundwork. The system’s approach to handling uncertainty, providing explanations, and structuring knowledge influenced countless AI applications, cementing MYCIN’s legacy as a cornerstone in the history of artificial intelligence in medicine.
References for Further Reading
Shortliffe, E. H. (1976). Computer-Based Medical Consultations: MYCIN. Elsevier.
Lindsay, Robert K., Bruce G. Buchanan, Edward A. Feigenbaum, and Joshua Lederberg. Applications of Artificial Intelligence for Organic Chemistry: The Dendral Project. McGraw-Hill Book Company, 1980.
Buchanan, B. G., & Shortliffe, E. H. (1984). Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison-Wesley.
Clancey, W. J., & Shortliffe, E. H. (1984). Readings in Medical Artificial Intelligence: The First Decade. Addison-Wesley.
5.3. CADUCEUS: A Successor to INTERNIST-I with Comprehensive Knowledge Base
CADUCEUS (Computer-Aided Deduction Using Experts) was an expert system developed in the early 1980s as an enhancement of the INTERNIST-I system. Designed mainly by Dr. Harry E. Pople Jr. and his colleagues at the University of Pittsburgh, CADUCEUS aimed to assist physicians in diagnosing complex internal medicine cases. Building upon the limitations of its predecessor, CADUCEUS incorporated more sophisticated reasoning methods and a comprehensive knowledge base to improve diagnostic accuracy. Its development marked a significant milestone in the application of artificial intelligence to healthcare, influencing subsequent computer-aided diagnosis systems throughout the 1980s and beyond.
CADUCEUS came about as a successor system to INTERNIST-I, created in response to the shortcomings identified in INTERNIST-I’s design. The INTERNIST-I program used a shallow causal graph and a sequential problem-solving approach to handle differential diagnosis based on observed symptoms. However, its limitations made it less acceptable to the medical community, even though it could still reach decisions.
To address these deficiencies, CADUCEUS was designed with a more advanced knowledge representation. It introduced multiple nosologic structures, allowing diseases to be classified in different ways, and incorporated a detailed pathophysiology represented by a causal graph with no restriction on its level of resolution. CADUCEUS also added generalized links to the causal graph, enabling rapid convergence on hypotheses, while still allowing detailed access to the underlying knowledge. This new design aimed to improve diagnostic accuracy and the system’s acceptance in the medical field.
How CADUCEUS Worked
Purpose
Diagnostic Assistance: Intended to support physicians in diagnosing a wide range of internal medicine disorders by analyzing patient data.
Educational Tool: Served as a learning resource for medical students and practitioners to understand diagnostic reasoning processes.
Knowledge Base
Comprehensive Medical Database: Contained detailed information on hundreds of diseases, symptoms, and clinical findings relevant to internal medicine.
Hierarchical Structure: Organized medical knowledge hierarchically, allowing the system to handle general and specific disease categories effectively.
Inference Engine
Modified Bayesian Reasoning: Utilized probabilistic reasoning to calculate the likelihood of various diagnoses based on patient data.
Heuristic Rules: Incorporated medical heuristics and empirical rules derived from expert clinicians to guide the diagnostic process.
User Interaction
Input Mechanism: Physicians entered patient information, including symptoms, medical history, laboratory results, and physical examination findings.
Output: Generated a ranked list of possible diagnoses along with explanations and justifications for each suggestion.
Technological Foundations
Programming Language
LISP Implementation: CADUCEUS was developed using LISP, leveraging its strengths in symbolic computation and rapid prototyping for AI applications.
Knowledge Representation
Semantic Networks: Employed semantic networks to represent medical concepts and their interrelationships, facilitating efficient retrieval and reasoning.
Frame-Based Structures: Used frames to encapsulate disease entities, symptoms, and findings, enabling modular and organized knowledge representation.
Computational Environment
Mainframe and Early Workstations: Operated on mainframe computers and early workstation systems, which provided the computational resources needed for complex processing.
Operational Workflow
Data Entry: The physician inputs detailed patient data, including all relevant clinical information and laboratory test results.
Initial Analysis: CADUCEUS analyzes the input data to identify salient features and preliminary diagnostic possibilities.
Hypothesis Generation: Generates a list of potential diagnoses ranked by their calculated probabilities.
Diagnostic Refinement: Allows the physician to provide additional data or clarify existing information to refine the diagnostic suggestions.
Final Output: Presents the most probable diagnoses along with explanations, supporting evidence, and suggested next steps for confirmation.
Technological Innovations
Advanced Probabilistic Reasoning: Improved upon earlier systems by implementing more sophisticated probabilistic models to handle diagnostic uncertainty.
Interactive Diagnostic Process: Featured an interactive dialogue with the user, allowing for dynamic refinement of diagnostic hypotheses.
Extensive Knowledge Base: Expanded the breadth and depth of medical knowledge compared to its predecessors, covering a wider range of diseases and conditions.
Explanatory Capabilities: Provided detailed explanations for its diagnostic suggestions, enhancing transparency and user trust.
Impact and Legacy
Enhanced Diagnostic Accuracy: Demonstrated improved performance in diagnosing complex cases, showcasing the potential of AI in enhancing clinical decision-making.
Educational Value: Used as a teaching tool in medical education, helping students understand the intricacies of diagnostic reasoning.
Influence on Future Systems: Informed the development of subsequent expert systems by highlighting the importance of probabilistic reasoning and user interaction.
Research Contributions: Contributed to academic research in artificial intelligence and medical informatics, providing insights into effective knowledge representation and reasoning strategies.
Challenges Highlighted: Brought attention to issues such as the complexity of knowledge base maintenance, integration with clinical workflows, and the need for user-friendly interfaces.
Operational Example
Consider a physician faced with a patient exhibiting symptoms like fatigue, joint pain, fever, and a rash. Using CADUCEUS, the physician inputs all clinical findings, including laboratory test results showing elevated inflammatory markers. CADUCEUS processes this information and generates a list of potential diagnoses such as systemic lupus erythematosus, rheumatoid arthritis, and infectious mononucleosis. It ranks these possibilities based on calculated probabilities and provides explanations for each, citing the supporting evidence from the patient’s data. The physician can then use this information to guide further testing and confirm the diagnosis.
Technological Advancements Inspired by CADUCEUS
Probabilistic Reasoning in AI: Emphasized the effectiveness of Bayesian and probabilistic methods in handling uncertainty in medical diagnoses.
User-Centered Design: Highlighted the importance of interactive and responsive user interfaces, influencing the design of future medical software to be more user-friendly.
Knowledge Base Management: Showcased the need for efficient methods to update and maintain large medical knowledge bases, leading to advancements in knowledge engineering practices.
Integration with Clinical Practice: Encouraged efforts to seamlessly integrate AI systems into clinical workflows, addressing practical challenges in healthcare settings.
Interdisciplinary Collaboration: Fostered greater collaboration between clinicians, computer scientists, and knowledge engineers, promoting interdisciplinary approaches in system development.
CADUCEUS stands as a significant advancement in the application of artificial intelligence to medicine during the 1980s. By enhancing diagnostic reasoning through advanced probabilistic models and interactive user engagement, it contributed substantially to the evolution of expert systems in healthcare. Its legacy is evident in the continued development of sophisticated diagnostic tools that build upon the foundations laid by early systems like CADUCEUS, aiming to improve patient outcomes through technology-enhanced clinical decision support.
References for Further Reading
Pople, H. E. (1982). Heuristic methods for imposing structure on ill-structured problems: The structuring of medical diagnostics. In P. Szolovits (Ed.), Artificial Intelligence in Medicine (pp. 119-190). Massachusetts Institute of Technology. https://people.csail.mit.edu/psz/ftp/AIM82/ch5.html
Miller, R. A., Pople, H. E. Jr., & Myers, J. D. (1982). Internist-I, an experimental computer-based diagnostic consultant for general internal medicine. New England Journal of Medicine, 307(8), 468-476.
Miller, R. A. (1984). Medical diagnostic decision support systems—past, present, and future: a threaded bibliography and brief commentary. Journal of the American Medical Informatics Association, 1(1), 8-27.
Miller, R. A. (1994). Evaluating evaluations of medical diagnostic systems. Journal of the American Medical Informatics Association, 1(1), 28-48.
5.4. QMR: A Superior Computational Knowledge Base Outpacing INTERNIST-I and CADUCEUS
The Quick Medical Reference (QMR) system was a significant advancement in medical artificial intelligence during the 1980s. Developed by Dr. Randolph A. Miller and his colleagues at the University of Pittsburgh in 1985, QMR was designed to assist physicians in diagnosing complex cases in internal medicine. Building upon the extensive knowledge base of the earlier INTERNIST-I system, QMR aimed to provide a practical and user-friendly diagnostic decision-support tool for clinicians.
The original goal of the INTERNIST-I project, initiated in the early 1970s, was to create an expert consultant program for diagnosing conditions in general internal medicine. By the early 1980s, it became clear that the project’s most valuable outcome was its medical knowledge base (KB), which was by then part of CADUCEUS. To leverage this KB in ways that benefited education, clinical practice, and computational analysis, the QMR program was developed. QMR provided a comprehensive summary of information from medical literature related to diagnosing disorders within internal medicine. Using common microcomputers, the program operated on three levels: as an electronic textbook, as a mid-level spreadsheet for combining and exploring basic diagnostic concepts, and as an expert consultant. The electronic textbook included an average of 85 findings and 8 associated disorders for diagnosing around 600 internal medicine disorders. By inverting these disease profiles, the system generated differential diagnosis lists for more than 4,250 patient findings. Unlike traditional medical textbooks, the QMR knowledge base allowed users to interact with and format displays dynamically to fit their specific information needs.
How QMR Worked
Purpose
Diagnostic Assistance: Aimed to help internists generate accurate differential diagnoses by analyzing patient symptoms, signs, and laboratory data.
Clinical Decision Support: Provided detailed information about diseases, aiding physicians in clinical decision-making and patient management.
Educational Resource: Served as a learning tool for medical students and residents, enhancing their understanding of disease presentations and diagnostic reasoning.
Knowledge Base
INTERNIST-I Legacy: Incorporated and expanded upon the INTERNIST-I knowledge base, which included detailed profiles of over 600 diseases in internal medicine.
Disease Profiles: Contained comprehensive information on diseases, including pathophysiology, typical and atypical manifestations, and associations with various clinical findings.
Manifestation Descriptions: Included over 4,000 clinical findings, with data on how frequently each finding was associated with specific diseases.
Inference Engine
Scoring Algorithm: Employed a sophisticated scoring system that calculated the likelihood of diseases based on the presence, absence, or negation of clinical findings.
Hypothesis Generation: Generated diagnostic hypotheses by matching patient data against disease profiles and assessing the strength of associations.
Heuristic Reasoning: Utilized heuristic methods to handle incomplete or uncertain data, enabling the system to function effectively even when patient information was limited.
User Interaction
Input Mechanism: Featured an interactive interface where physicians could enter patient data using natural language or coded inputs for symptoms, signs, and test results.
Query System: Allowed users to ask questions about diseases, manifestations, and the reasoning behind diagnostic suggestions.
Output: Provided a ranked list of potential diagnoses, detailed explanations of each disease, suggested additional tests, and rationale for its conclusions.
Technological Foundations
Programming Language
LISP: QMR utilized LISP as its programming language, with the diagnostic algorithms being specifically implemented in LISP-based structures, allowing for efficient processing of medical data within the system.
Knowledge Representation
Frame-Based Architecture: Used frames to represent diseases and clinical findings, organizing knowledge into structured data entities.
Probabilistic Associations: Incorporated statistical data on disease prevalence and the likelihood of clinical findings, enhancing diagnostic accuracy.
Hierarchical Structuring: Organized diseases and findings hierarchically, allowing for efficient retrieval and reasoning processes.
Computational Environment
Personal Computers: Designed to run on IBM-compatible personal computers, making it accessible to a broader range of healthcare providers without the need for specialized hardware.
Efficient Algorithms: Optimized computational algorithms to function effectively within the limited processing power and memory constraints of PCs at the time.
Operational Workflow
Data Entry: The physician inputs detailed patient information, including symptoms, physical examination findings, laboratory results, and imaging studies.
Data Processing: QMR processes the input data, mapping findings to its knowledge base and identifying relevant disease entities.
Hypothesis Generation: Generates a list of potential diagnoses by comparing patient data with disease profiles and calculating diagnostic scores.
Diagnostic Ranking: Ranks the potential diagnoses based on their computed scores, indicating the most probable conditions.
Information Retrieval: Provides detailed information on each suggested disease, including typical presentations, associated findings, and recommended diagnostic tests.
Iterative Refinement: Allows physicians to input additional data or modify existing information to refine the diagnostic suggestions.
Technological Innovations
User-Centric Design: Emphasized ease of use, with an interface designed to fit seamlessly into clinical workflows and minimize disruption.
Integration of Comprehensive Knowledge Base: Combined an extensive medical knowledge base with practical decision-support capabilities.
Advanced Scoring Mechanisms: Developed sophisticated algorithms to handle complex diagnostic reasoning, incorporating both probabilistic and heuristic methods.
Adaptability: Designed to handle a wide range of cases in internal medicine, from common conditions to rare diseases, making it a versatile tool for clinicians.
Impact and Legacy
Clinical Acceptance: Achieved a higher level of acceptance among clinicians compared to earlier systems due to its practicality and relevance to daily practice.
Educational Impact: Used extensively in medical education to teach diagnostic reasoning and the relationships between diseases and clinical findings.
Foundation for Future Systems: Influenced the development of subsequent diagnostic and decision-support systems, emphasizing the importance of usability and integration with clinical practice.
Research Contributions: Provided valuable insights into the challenges of knowledge representation, reasoning under uncertainty, and the practical application of AI in medicine.
Highlighting Limitations: Exposed limitations related to the maintenance and updating of large medical knowledge bases, spurring research into more scalable solutions.
Operational Example
Imagine a physician evaluating a patient who presents with symptoms such as persistent cough, weight loss, night sweats, and enlarged lymph nodes. Concerned about a wide range of possible conditions, the physician inputs these findings into QMR. The system processes the information and generates a differential diagnosis list that includes diseases such as tuberculosis, lymphoma, and sarcoidosis, ranked according to their likelihood based on the input data. QMR provides detailed information on each disease, including typical and atypical features, suggested diagnostic tests, and potential management strategies. This assists the physician in considering diagnoses that may not have been immediately apparent and guides further investigation and treatment planning.
Technological Advancements Inspired by QMR
Enhanced Knowledge Representation: Highlighted the need for more efficient methods of knowledge representation, leading to the development of object-oriented and ontology-based models in medical AI.
Evidence-Based Decision Support: Inspired the integration of evidence-based medicine principles into decision-support systems, emphasizing the use of up-to-date clinical guidelines and research findings.
User Interface Design: Influenced the design of more intuitive and interactive user interfaces, incorporating graphical elements and natural language processing to improve user experience.
Interoperability and Integration: Encouraged the development of systems capable of integrating with electronic health records (EHRs) and other clinical information systems, enhancing data sharing and workflow efficiency.
Research in Uncertainty Management: Stimulated further research into probabilistic reasoning and Bayesian methods to better handle uncertainty in medical diagnosis.
QMR stands as a landmark in the evolution of medical diagnostic systems, bridging the gap between theoretical AI applications and practical clinical tools. By focusing on usability, comprehensive knowledge, and effective reasoning methods, QMR demonstrated the potential for AI to enhance diagnostic accuracy and support clinicians in managing complex medical cases. Its legacy continues to influence the design and implementation of modern decision-support systems, contributing to ongoing advancements in medical informatics and artificial intelligence.
References for Further Reading
Miller, R. A., Pople, H. E., & Myers, J. D. (1982). INTERNIST-1, an experimental computer-based diagnostic consultant for general internal medicine. New England Journal of Medicine, 307(8), 468–476.
Miller, R. A., Masarie, F. E., & Myers, J. D. (1986). QUICK MEDICAL REFERENCE (QMR) for diagnostic assistance. MD Computing, 3(5), 34–48.
Masarie, F. E., & Miller, R. A. (1988). Medical subject headings (MeSH) as a knowledge source for clinical decision support systems. Proceedings of the Annual Symposium on Computer Application in Medical Care, 101–105.
Pople H. E. The QMR computer project in perspective: Reflections on a decade of medical artificial intelligence. MD Comput. 1989 May-Jun;6(3):282-9. PMID: 2695783.
Miller, R. A. (1994). Medical diagnostic decision support systems—past, present, and future: a threaded bibliography and brief commentary. Journal of the American Medical Informatics Association, 1(1), 8–27.
5.5. CADIAG: The Birth of Fuzzy Logic in Medical Diagnosis
CADIAG, short for Computer-Aided Diagnosis, is a series of medical expert systems developed to assist physicians in the diagnostic process. Developed in the late 1970s by Klaus-Peter Adlassnig and his colleagues at the University of Vienna and the Ludwig Boltzmann Institute for Rheumatology and Focal Diseases, CADIAG aims to improve diagnostic accuracy by utilizing artificial intelligence techniques, such as fuzzy logic and probabilistic reasoning.
CADIAG-1 was based on symbolic logic and relied on predefined strong and weak relationships between symptoms and diseases. Strong relationships include confirming or excluding diagnoses, while weak relationships, termed FN-relationships (facultative and not confirming), were used for creating diagnostic hypotheses by combining symptoms. This system was built around more rigid, rule-based logic, primarily using binary (true or false) logic for medical relationships.
CADIAG-2, however, was developed using fuzzy set theory and fuzzy logic, which allowed for a more flexible and nuanced representation of medical relationships. CADIAG-2 introduced the ability to handle imprecise or vague data by assigning degrees of membership to different symptoms and diagnoses. Instead of binary relationships, it incorporated a range of values, reflecting how strongly a symptom might suggest a particular diagnosis. This enabled a more detailed analysis of complex medical cases, especially where symptoms and diseases might overlap or be unclear.
The idea of Fuzzy Logic originated from the work of Lotfi A. Zadeh, “the father of Fuzzy Logic” and a professor at the University of California, Berkeley. He introduced the concept in his seminal paper titled “Fuzzy Sets” in 1965. Zadeh proposed Fuzzy Logic as a way to model and reason with uncertainty and imprecision, which traditional Boolean logic (true/false, yes/no) could not adequately handle.
In layman’s terms, Fuzzy Logic is a precise tool for dealing with the imprecise real world.
In contrast to classical logic, where variables are either fully true or fully false, Fuzzy Logic allows for a degree of truth, meaning that variables can take any value between 0 (completely false) and 1 (completely true). This was particularly useful in areas like control systems, pattern recognition, decision-making, and, as seen with CADIAG, medical diagnosis, where many real-world situations involve vague or partial information.
Zadeh’s work was revolutionary because it enabled more flexible reasoning, mimicking how humans make decisions based on incomplete or ambiguous data, which had vast applications in various fields, especially artificial intelligence, engineering, and medicine.
In 1985, Masaki and Watanabe at Bell Labs developed the first fuzzy logic chip.
In fact, fuzzy logic became an integral part of Bell Labs’ artificial intelligence and machine learning research as it allowed systems to process data in a manner that mimicked human reasoning. The ability to handle partial truths enabled more intelligent systems that could learn, adapt, and make decisions based on complex, multi-variable inputs. This laid some of the groundwork for advancements in expert systems and other AI technologies.
How CADIAG Worked
Purpose
Diagnostic Assistance: Aimed to support physicians in the differential diagnosis process by analyzing patient symptoms, clinical findings, and laboratory results.
Decision Support: Provided probabilistic assessments of potential diagnoses, helping clinicians consider a range of possible conditions based on varying degrees of certainty.
Knowledge Base
Medical Knowledge Representation: Contained an extensive database of diseases, symptoms, signs, and laboratory findings specific to internal medicine.
Fuzzy Relationships: Utilized fuzzy set theory to represent the degrees of association between clinical findings and diseases, acknowledging the non-binary nature of medical information.
Inference Engine
Fuzzy Logic Reasoning: Employed fuzzy logic to process imprecise and uncertain data, allowing for partial truth values rather than binary true/false outcomes.
Probabilistic Assessments: Calculated degrees of diagnostic certainty by aggregating fuzzy relationships, enhancing the system’s ability to handle complex medical scenarios.
User Interaction
Input Mechanism: Clinicians entered patient data through an interface, including symptoms, physical examination results, and laboratory test outcomes.
Output: Generated a ranked list of possible diagnoses with associated certainty levels and provided explanations linking patient data to each suggested diagnosis.
Technological Foundations
Programming Language
PROLOG Implementation: CADIAG was implemented using PROLOG, a logic programming language well-suited for symbolic reasoning and handling complex data structures in artificial intelligence applications.
Knowledge Representation
Fuzzy Set Theory: Central to CADIAG’s knowledge representation, enabling the modeling of uncertain and imprecise relationships between medical findings and diseases.
Semantic Networks: Used to structure medical concepts and their interrelationships, facilitating efficient retrieval and reasoning processes.
Computational Environment
Mainframe and Early Personal Computers: Initially ran on mainframe systems due to the computational demands of fuzzy logic processing but was later adapted for early personal computers, increasing accessibility for medical practitioners.
Operational Workflow
Data Entry: The physician inputs comprehensive patient data, including symptoms, signs, and laboratory findings.
Fuzzy Matching: CADIAG compares the input data against its knowledge base, assessing the degree of match between patient data and potential diagnoses using fuzzy logic.
Diagnostic Reasoning: Aggregates the fuzzy relationships to calculate degrees of certainty for possible diagnoses.
Output Generation: Presents a ranked list of potential diagnoses with associated certainty levels and explanatory information.
Feedback Loop: Allows physicians to provide feedback or additional data, refining the system’s recommendations in subsequent iterations.
Technological Innovations
Integration of Fuzzy Logic: Pioneered the use of fuzzy logic in medical diagnosis, allowing for nuanced handling of uncertain and imprecise information.
Advanced Knowledge Representation: Developed sophisticated methods for representing medical knowledge, including the use of semantic networks and fuzzy relationships.
Probabilistic Reasoning: Enhanced the decision-making process by providing probabilistic assessments rather than deterministic outcomes.
User-Centric Design: Focused on creating an interface that accommodated the needs of physicians, facilitating easier adoption in clinical settings.
Impact and Legacy
Advancement of Medical AI: CADIAG’s successful application of fuzzy logic influenced subsequent medical expert systems and demonstrated the viability of AI in clinical diagnostics.
Educational Contributions: Served as a foundation for academic research and education in medical informatics and artificial intelligence.
Interdisciplinary Collaboration: Fostered collaboration between medical professionals and computer scientists, promoting interdisciplinary approaches to healthcare challenges.
Improved Diagnostic Processes: Aided in complex diagnostic scenarios, potentially improving patient outcomes by assisting in early and accurate disease identification.
Operational Example
Imagine a patient presenting with symptoms like fatigue, joint pain, and elevated inflammatory markers. The physician inputs these symptoms and laboratory findings into CADIAG. The system uses fuzzy logic to evaluate the degree of association between the patient’s data and various rheumatologic conditions, such as rheumatoid arthritis or systemic lupus erythematosus. CADIAG then provides a ranked list of potential diagnoses with corresponding certainty levels, offering the physician valuable insights that might prompt further specific testing or immediate intervention.
Technological Advancements Inspired by CADIAG
Enhanced Decision Support Systems: Influenced the development of more sophisticated medical decision support tools capable of handling complex and uncertain data.
Application of Fuzzy Logic in AI: Demonstrated the effectiveness of fuzzy logic in artificial intelligence, leading to its adoption in various domains beyond medicine.
Improved Knowledge Engineering: Advanced techniques in knowledge representation and reasoning, contributing to the evolution of expert systems.
CADIAG stands as a significant milestone in the application of artificial intelligence to medicine. Its innovative use of fuzzy logic to handle the complexities and uncertainties of medical diagnosis paved the way for more advanced expert systems. By enhancing diagnostic accuracy and efficiency, CADIAG contributed to improving patient care and demonstrated the profound potential of AI technologies in healthcare. Its legacy continues to influence the development of intelligent medical systems and the integration of AI into clinical practice.
References for Further Reading
Adlassnig, K.-P. (1980). A Fuzzy Logical Model of Computer-Assisted Medical Diagnosis. Methods of Information in Medicine, 19(3), 141-148.
Adlassnig, K.-P., & Kolarz, G. (1982). CADIAG-2: Computer-Assisted Medical Diagnosis Using Fuzzy Subsets. In M. Gupta & E. Sanchez (Eds.), Approximate Reasoning in Decision Analysis (pp. 219-229). North-Holland Publishing Company.
Adlassnig, K.-P., Kolarz, G., Scheithauer, W., Effenberger, H., & Grabner, G. (1985). CADIAG: Approaches to computer-assisted medical diagnosis. Computers in Biology and Medicine, 15(5), 315-335. https://doi.org/10.1016/0010-4825(85)90014-9.
Adlassnig, K.-P., & Kolarz, G. (1986). Representation and Semantics of Fuzzy Medical Knowledge in CADIAG-1 and CADIAG-2. Computers and Biomedical Research, 19(1), 63-79.
Adlassnig, K.-P., Kolarz, G., Scheithauer, W., & Grabner, G. (1985). Approach to a Hospital-Based Application of a Medical Expert System. Medical Informatics, 10(3), 205-223.
Klinov, P., Parsia, B., & Picado-Muiño, D. (2010). The Consistency of the CADIAG-2 Knowledge Base: A Probabilistic Approach. In P. Grzegorzewski et al. (Eds.), Advances in Soft Computing and Its Applications: 9th International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, IPMU 2010 (pp. 314-325). Springer.
5.6. DXplain: The Pioneer of Computer-Based Diagnostic Decision Support
DXplain is a clinical decision support system developed in the mid-1980s at the Laboratory of Computer Science, Massachusetts General Hospital. Conceived by Dr. G. Octo Barnett and his colleagues, DXplain was designed to assist clinicians in generating differential diagnoses based on patient-specific clinical findings. Serving both as a diagnostic aid and an educational tool, DXplain has significantly influenced the field of medical informatics and the development of subsequent expert systems in healthcare.
The American Medical Association supported and cooperated in the development of DXplain.
Dr. Barnett is also known for co-developing COSTAR, one of the first computerized electronic health records. He is considered a founding father of medical informatics.
In terms of integration, DXplain had been incorporated into various Clinical Information Systems, including the one at Columbia Presbyterian Medical Center, as described in the research. It could be accessed via web-based interfaces or through TCP/IP function calls, facilitating its use in a variety of healthcare settings.
How DXplain Worked
Purpose
Diagnostic Assistance: Aimed to help clinicians, who don’t have computer expertise, generate a list of possible diagnoses by analyzing a patient’s clinical manifestations, including symptoms, physical signs, and laboratory results.
Educational Tool: Served as a resource for medical education by providing detailed explanations of diseases and their associated clinical findings.
Knowledge Base
Extensive Medical Database: Contained information on over 2,000 diseases and thousands of clinical findings, capturing the associations between them.
Association Rules: Represented the relationships between diseases and clinical manifestations using probabilistic and heuristic associations.
Inference Engine
Probabilistic Reasoning: Employed statistical methods to calculate the likelihood of various diagnoses based on the inputted clinical findings.
Differential Diagnosis Generation: Produced a ranked list of potential diagnoses, considering both the presence and absence of specific clinical findings.
User Interaction
Input Mechanism: Users entered patient data through an interface, selecting clinical findings from predefined lists or inputting them manually.
Output: Provided a differential diagnosis list, detailed explanations for each suggested disease, and the rationale behind their inclusion.
Technological Foundations
Programming Language
MUMPS: Initially developed using the Massachusetts General Hospital Utility Multi-Programming System (MUMPS), a language optimized for handling medical data.
Knowledge Representation
Statistical Associations: Modeled the diagnostic reasoning process through statistical relationships between diseases and clinical findings.
Heuristic Rules: Incorporated heuristic knowledge to refine diagnostic suggestions and manage exceptions.
Computational Environment
Mainframes to Microcomputers: Originally ran on hospital mainframe systems and was later adapted for microcomputers and online platforms, enhancing accessibility.
Operational Workflow
Data Entry: The clinician inputs patient-specific clinical findings, including symptoms, physical examination results, and laboratory data.
Data Analysis: DXplain analyzes the inputted findings, matching them against its knowledge base to identify diseases that could explain the observed findings.
Differential Diagnosis Generation: Produces a ranked list of potential diagnoses, prioritizing them based on how well they match the patient’s clinical presentation.
Explanation and Justification: Provides detailed explanations for each suggested diagnosis, including reasons for their inclusion and pertinent clinical information.
Educational Feedback: Offers additional learning resources, such as disease descriptions and management guidelines.
Technological Innovations
Large-Scale Knowledge Base: One of the first systems to incorporate an extensive database covering a broad range of diseases and clinical findings.
Interactive Diagnostic Reasoning: Allowed users to iteratively refine input data and observe how changes affected the differential diagnosis.
Educational Integration: Combined diagnostic assistance with educational content, enhancing its utility as a teaching tool.
User-Friendly Interface: Designed to be accessible to clinicians with varying levels of computer proficiency, facilitating wider adoption.
Impact and Legacy
Widespread Adoption: Used by numerous medical institutions and educational programs, contributing to improved diagnostic accuracy and medical education.
Influence on Medical Informatics: Served as a model for subsequent clinical decision support systems, demonstrating practical AI applications in medicine.
Research Contributions: Generated valuable insights into diagnostic reasoning processes, informing future research in medical expert systems.
Continued Evolution: Updated over time with enhancements to its knowledge base, interface, and platform compatibility, remaining relevant amid technological advancements.
Operational Example
Imagine a physician evaluating a patient who presents with fever, rash, and joint pain. The clinician inputs these clinical findings into DXplain. The system analyzes the data and generates a differential diagnosis list that includes conditions like systemic lupus erythematosus, rheumatoid arthritis, and Lyme disease. For each suggested diagnosis, DXplain provides explanations, typical clinical features, relevant tests, and management considerations. This assists the physician in considering possible diagnoses and guides further diagnostic workup.
Technological Advancements Inspired by DXplain
Enhanced Knowledge Representation: Influenced the development of sophisticated methods for representing medical knowledge, including ontologies and semantic networks.
Integration with Electronic Health Records: Paved the way for integrating decision support systems with electronic health records, enabling real-time diagnostic assistance.
Mobile and Web-Based Platforms: Inspired the adaptation of clinical decision support tools to mobile and web-based platforms, increasing accessibility.
Patient Safety Initiatives: Contributed to efforts aimed at reducing diagnostic errors and improving patient safety through enhanced clinical decision support.
DXplain stands as a significant milestone in the application of artificial intelligence to clinical medicine. Developed in the 1980s, it showcased the potential of computer-assisted diagnostic reasoning and provided valuable educational resources. By facilitating the generation of differential diagnoses and offering detailed explanations, DXplain enhanced clinical decision-making and influenced the development of subsequent medical expert systems. Its legacy continues as it adapts to new technologies and remains a relevant tool in today’s healthcare informatics landscape.
References for Further Reading
Barnett, G. O., Cimino, J. J., Hupp, J. A., & Hoffer, E. P. (1987). DXplain: An evolving diagnostic decision-support system. JAMA, 258(1), 67-74.
Feldman MJ, Barnett GO. An approach to evaluating the accuracy of DXplain. Comput Methods Programs Biomed. 1991 Aug;35(4):261-6. doi: 10.1016/0169-2607(91)90004-d. PMID: 1752121.
Miller, R. A. (1994). Medical diagnostic decision support systems—past, present, and future: a threaded bibliography and brief commentary. Journal of the American Medical Informatics Association, 1(1), 8-27.
Elhanan, G., Socratous, S. A., & Cimino, J. J. (1996). Integrating DXplain into a Clinical Information System using the World Wide Web. Proceedings of the AMIA Annual Fall Symposium, 347-351.
Bartold, S. P., & Hannigan, G. G. (2002). Software review: DXplain. Journal of the Medical Library Association, 90(2), 267-268.
Berner, E. S., & La Lande, T. J. (2007). Overview of clinical decision support systems. In Clinical Decision Support Systems (pp. 3-22). Springer.
That wraps up Part 1. In Part 2, we’ll dig into why so many people are convinced AI will soon replace doctors. Exciting? Yes. Scary? Absolutely, if it’s true.
Stay tuned…
Hi! I believe in fair compensation for every piece of work. So, to those who’ve stepped up as paid subscribers: thank you. 🙏 I don’t have sponsors. I don’t have advertisers. I don’t have bosses. Why? Because I want my research and analyses to remain original and unbiased. However, to continue my work, I rely on support from readers like you. If you’re not a paid subscriber yet and you want my relentless research and analyses like this one to continue for another 12 months, 24 months, 36 months, and more, consider joining the ranks:
👉👉👉👉👉 My name is Sergei Polevikov. I’m an AI researcher and a healthcare AI startup founder. In my newsletter ‘AI Health Uncut’, I combine my knowledge of AI models with my unique skills in analyzing the financial health of digital health companies. Why “Uncut”? Because I never sugarcoat or filter the hard truth. I don’t play games, I don’t work for anyone, and therefore, with your support, I produce the most original, the most unbiased, the most unapologetic research in AI, innovation, and healthcare. Thank you for your support of my work. You’re part of a vibrant community of healthcare AI enthusiasts! Your engagement matters. 🙏🙏🙏🙏🙏