Skip to main content

History of AI/Medicine by Ted Shortliffe

Overview​

Today’s state of the art in medical AI is part of a 60-year scientific trajectory and is still evolving.

AI in image generation has historically been imperfect. Harold Cohen created AARON, which had exhibitions including a Whitney Museum exhibit.

Origins of AI​

  • “Artificial Intelligence” coined by John McCarthy
  • 1959: first computer to play checkers
  • Key leaders in the 1960s: Herb Simon, Allen Newell, Marvin Minsky, John McCarthy
  • 1960–1970: substantial AI work on problems such as ML/neural networks, NLP, speech understanding, simulation of human problem solving, and early robotics

AI as Described in 1982​

  • The ability to reason symbolically
  • The ability to acquire and apply knowledge
  • The ability to manipulate and communicate ideas

Early AI and Biomedicine​

The Dendral Project: scientific hypothesis formation and discovery; encoded knowledge of organic chemistry.

Example Projects​

  • CASNET: causal association network
  • MYCIN: backward chaining expert system
  • Internist-1: evoking strengths, simple “naive” Bayesian models, certainty factors

A foundational insight from this era: clinicians will use AI systems if the programs can be shown to function at the level of experts.

The AI Hype Cycles​

  • Explosive interest in AI beyond academia
  • 1980s–90s: disappointment that applied AI had not rapidly lived up to predictions; perceived overselling of AI
  • Periods of disillusionment came to be called “AI winters”

Evolution in Technology, Methods, and Applications​

  • ARPAnet generalized and gradually became commercial (the Internet), with the introduction of the domain system in the late 1980s
  • Introduction of PCs (1980)
  • Shrinking size and cost of computer memory
  • Rapidly increasing computational power, introduction of GPUs
  • Broadening of application domains
  • Large databases
  • Other AI opportunities: terminology, ontologies, LLMs, NLP

Fascination with the role of technology and computing in health care persisted.

Transition to the EHR​

  • Large amounts of digital patient data
  • New capabilities due to visionary work of computer scientists such as Geoffrey Hinton
  • Huge opportunity to learn from millions of cases

Applications in Clinical Practice​

Medical Device Data Interpretation​

Decision-specific clinical data → software produces interpretive report → delivered to clinician → clinician verifies interpretation and incorporates it into the decision-making process.

Event Monitoring and Alerts​

Repository for clinical information on a patient → event monitoring software ← knowledge base on what to watch.

Direct Consultation with Clinical User​

Clinical advisory tools draw from:

  • Analytical methods
  • Repository for clinical information on a patient

These feed into an interface to the clinician that integrates decision support with chart review and order entry.

Explainable AI​

ML models are opaque, non-intuitive, and difficult for people to understand. Demonstrably accurate analysis may lead to trust of AI models.

Criteria for Decision Support Acceptance and Integration into Workflow​

  • Efficient
  • Intuitive
  • Simple to learn
  • Reflective and understanding of the pertinent domain

Open Question​

Is an algorithm derived using deep learning methods new knowledge?

Lecture Summary​

The lecture offered a history of AI and its application to healthcare and medicine. The term “Artificial Intelligence” was coined by John McCarthy (he of LISP and Lambda Calculus fame) in the late 1950s. He coined this with another giant, Claude Shannon, for a conference at Dartmouth. Other pioneers are Herb Simon, Marvin Minsky, and Allen Newell.

Through the 60s and 70s, there was a lot of foundational and ground-breaking work done on things we know today as NLP, speech processing, neural networks, and machine learning (the first “artificial neuron” was first described in the 1940s, however, and the Perceptron was developed in the late 1950s).

By the 1980s, “AI” was generally understood to be a machine/system to which many definitions of “intelligence” could be applied: acquire and apply knowledge (this is a dictionary definition), manipulate and communicate information/ideas, reason in a symbolic way. However, this decade and the next led to the first “AI Winter,” where all the hype and promise of AI sort of fizzled away.

(Some discussion on early efforts at applying AI to healthcare.)

Our times are unlike these, however, due to vast improvements in computing power (Moore’s Law, GPUs) and the digitization of health records. This has led to new areas of inquiry and application like clinical decision systems.

Discussion on Explainable AI and its importance (e.g., clinicians will “trust” AI if it can be shown to operate at expert levels). Opaque models help nobody (except VCs/financebros and techbros).