AI, BIOMEDICINE, AND THE NIH 1st ACM International Health Informatics Symposium 11/11/2010 Milton Corn, M.D., National Library Of Medicine National Institutes Of Health

What is AI? —  —  —  — 

— 

CIRCULAR: Artificial intelligence (AI) is the intelligence of machines GLOBAL: Understanding the mechanisms underlying thought and intelligent behavior and their embodiment in machines. SIMPLISTIC: Software that passes the Turing test PRAGMATIC: —  machines that perform functions that require intelligence when performed by humans —  artificial systems that will perform better on tasks that humans currently do better e.g. in expert systems, speech recognition, machine vision, searching, speech recognition, reading METHODOLOGIC: solving problems with expert systems, neural nets, machine learning, genetic algorithms

IS DEEP BLUE AI?

Cousins of AI —  computational intelligence —  machine learning —  intelligence amplifying systems —  flexible competence —  human-computer collaboration —  computational thinking

Hollywood Knows AI

Birth Of AI In Modern Era —  1956 Dartmouth Conference —  Organizers: Marvin Minsky, John McCarthy and two IBM

scientists, Claude Shannon and Nathan Rochester —  Conference assertion: "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it".

Promises: —  1958, H. A. Simon and Allen Newell: "within ten years a

computer will be the world's chess champion" and "within ten years a digital computer will discover and prove an important new mathematical theorem.” —  1965, H. A. Simon: "machines will be capable, within twenty years, of doing any work a man can do.” —  1967, Marvin Minsky: "Within a generation ... the problem of creating 'artificial intelligence' will substantially be solved.” —  1970, Marvin Minsky (in Life Magazine): "In from three to eight years we will have a machine with the general intelligence of an average human being.”

Seasons of AI

Gartner Inc.'s Hype Cycle showing the fluctuating support of a new technology

SOME AI ACHIEVEMENTS —  —  —  —  —  —  —  —  —  — 

machine translation data mining industrial robotics, speech recognition banking software medical diagnosis, Google's search engine fuzzy logic controllers heuristic search data analytics

WHAT IS GOING ON NOW? Some Important Current AI Projects —  Personal assistants (e.g. Microsoft) —  Speech recognition (e.g. Siri) —  “Watson” for question-answering (IBM) —  Robo-docs (e.g. X Prize $10 Million For AI Diagnostician) —  NELL: Never-Ending Language Learning machine that

“reads” (Carnegie Mellon) —  DARPA Programs e.g Grand Challenges, Semantic Web —  EU-FP7 Support

HOW DID AI INTERACT WITH MEDICINE? —  Analogous “hype” cycle over the decades —  Early emphasis on “doc in the box” e.g. diagnosis programs —  Heavy use of rules-based expert systems

Some early focused programs —  MYCIN: rules-based system, developed at Stanford originally

for the diagnosis and treatment of bacterial infections of the blood and later extended to handle other infectious diseases as well. —  Present Illness Program (PIP), for taking the history of the present illness of a patient with renal (kidney) disease —  ABEL, a program for the diagnosis (and eventually treatment) of acid/base and electrolyte disturbances.

Some General Diagnosis Programs —  Iliad: Bayesian general diagnosis system that grew out of an

early electronic record, the HELP system —  INTERNIST-I a computer-assisted diagnostic tool designed to capture the expertise of just one man, Jack D. Myers, MD —  Dxplain uses modified Bayes to produce a ranked list of diagnoses, provide justification for each disease, suggest further steps

Some Early Lab Programs —  PUFF Expert System for Interpretation of Pulmonary Function

Data) was probably the first AI system to be used in clinical practice.   —  APACHE was one of the first medical decision support systems to be commercialized - Provides a numerical indicator of severity of illness, risk of mortality, risk of active treatment (an indication of appropriate resource utilization) to predict clinical results in the ICU —  GermWatcher checks for hospital-acquired (nosocomial) infections —  PEIRS (Pathology Expert Interpretative Reporting System) PEIRS interpreted about 80-100 reports a day with a diagnostic accuracy of about 95%

WERE AI PROGRAMS WELCOMED BY HEALTHCARE? —  Not by clinicians —  Not by educators —  Some for clinical surveillance —  Commonly in laboratories —  Significantly in signal analysis (ECG, EEG, EMG)

*WHY ARE AI PROGRAMS REJECTED?* —  Product not wanted or puerile —  Onerous, time-consuming data input —  Clumsy interface —  Program intrusive (esp for reminders) —  Plethora of single-issue programs —  Even small error rate not acceptable —  Inventor makes no effort to disseminate —  Not interfaced with EHR —  Not invented here —  Technophobia

WHAT ABOUT AI AND HEALTHCARE TODAY? Dep’t of Extravagant Predictions

Stead et al, Academic Medicine 86:1-6, 2011 —  List patients in order of acuity based on continuing analysis of

EHRs —  Propose differential diagnosis based on literature and computational biological model of patient —  Display costs of alternate management plans with explanation —  Base dx/rx on local disease prevalence/experience —  Discuss whether recommended management differs from that of general community

AI AND HEALTHCARE TODAY Topics for which funding available —  Decision support (CDDS) —  Generating alerts and reminders —  Diagnostic assistance —  Therapy critiquing and planning —  Guidelines —  HCI/Behavioral aspects of CDDS

—  Agents for information retrieval —  Image recognition and interpretation

WHAT SORT OF PROJECTS ARE BEING PUBLISHED? Artificial Intelligence in Medicine Vol 50, issue 2, Oct 2010 RESEARCH ARTICLES — 

Semantic relations for problem-oriented medical records

— 

Detecting ‘wrong blood in tube’ errors: Evaluation of a Bayesian network approach

— 

—  — 

Automatic image-based assessment of lesion development during hemangioma follow-up examinations Decision support in heart failure through processing of electro- and echocardiograms Missing data imputation using statistical and machine learning methods in a real breast cancer problem

— 

Evolving fuzzy medical diagnosis of Pima Indians diabetes and of dermatological diseases

— 

Quantitative prediction of MHC-II binding affinity using particle swarm optimization

THE TWO WORLDS OF AI Grand Challenges

Feasible Miniature Projects

—  Personal Assistants

—  Wrong blood in tube

—  Watson

—  Dx diabetes Pima Indians

—  Nell

—  Missing data in breast ca pt.

USUALLY:

USUALLY:

Private sector funding Profit potential

Federal funding

Dissemination likely

Promotion potential Dissemination rare

DOES NIH CARE ABOUT HEALTH-CARE AI? —  NLM funds healthcare informatics, including AI —  (mostly feasible miniatures through standard NIH grant

programs) —  NIBIB might be receptive to image , bio-engineering, or

robotics projects —  Other Institutes not focused on health-care delivery but —  Could be interested in AI for clinical research —  Could be interested in secondary use of EHR data —  Are interested in computational solutions to genome-related

problems

Why Is NLM Interested in AI? NLM MISSION IS ABOUT BIOMEDICAL INFORMATION Task 1: store and archive information Task 2: make the information accessible Task 3: facilitate utilization of information Premise: vast and rapidly expanding stores of biomedical information can be exploited more efficiently if human cognition is extended with AI tools

NLM AI CHALLENGE GRANTS & AI 2009-10 Challenge Grants Using Federal Stimulus Funds —  Intelligent search tool for answering clinical questions —  Self-documenting encounters between clinician and

patient or research subject —  Informatics for post-marketing surveillance —  Presenting individual genome information in patientrelevant, decision-supporting form —  Computational hypothesis generation for biology and medicine by analysis of databases —  In silico hypothesis testing for biology and medicine —  Create intelligent summary of an electronic health record

NLM RFP For Computational Thinking Spring-Summer 2010 —  Propose creative hypotheses based on an interrogation of

heterogeneous (experimental or computational) data sets. —  Language understanding of medical records —  Interpretation and intelligent summarization/visualization of data, text, or images —  Decision support under uncertainty with associated explanations —  Intelligent retrieval —  Programs that can explain their output —  13 AWARDS MADE; TO BE ANNOUNCED SHORTLY

Program To Read, Explain, Answer Questions About Rx Label, Side-effects, And Picture Of Tablet/Capsule

Applying To NIH —  Monitor NIH Guide for Announcements —  Consider needs of each Institute —  Address a problem (AI is an approach, not the goal) —  Try email to judge interest before writing entire application —  Have clinical expertise on your team —  Get help from NIH veteran if you are novice

For Further Information —  [email protected] —  NLM EXTRAMURAL PROGRAMS AND GRANTS

http://www.nlm.nih.gov/ep/index.html •  KEY SITE FOR NIH COMPUTING ACTIVITIES

http://www.bisti.nih.gov/ •  SITE FOR SMALL BUSINESS GRANTS

http://grants.nih.gov/grants/funding/sbir.htm