Keynote lecture delivered on at the NIPS workshop on health applications https://ml4health.github.io/2017/ in which I describe which AI applications really will make a health impact which may not be the same as those that deliver high ROI in the near term.
5. DEPARTMENT OF
Biomedical Informatics
A common diagnosis becomes undiagnosed
• 3 years and 10 months old with bloody stool.
• Underwent endoscopy in November 2005 at 4 years old which showed a pancolitis.
• He was started on sulfasalazine; maintained on fish oil and sulfasalazine until age 13 ½ at which
point he flared. That was in July 2015.
• He’s been in a flare since. Unsuccessful wean from prednisone, vancomycin effect transient
• Started on 6MP in March 2016 with no effect over 3 months despite therapeutic levels.
• Started on infliximab 5mg/kg 7/6/2016 with no effect.
• Aug 20th tried a course of Rifaximin with initial, temporally related, transient improvement.
• Has tried multiple forms of PR meds with no effect, including cortisone enemas, cortifoam,
canasa suppos. .
• In September 2016, stool turned bloody and frequency was hourly, hospitalized for tacrolimus
• which improved symptoms but did not produce a remission.
• Vedolizumab was added in October with no appreciable effect.
6. DEPARTMENT OF
Biomedical Informatics
Boundary between diseased and
healthy patients
Can we identify an existing drug that
will move these patients towards the
healthy region?
IBD Expression Profiles:
Whole Blood
8. DEPARTMENT OF
Biomedical Informatics
Best ranked compound for our patient
• Indirubin
• Chemical compound
most often produced
as a byproduct of
bacterial metabolism
• Constituent of indigo
naturalis (also known
as qing dai),
12. DEPARTMENT OF
Biomedical Informatics
When does Medicine succumb to AlphaZero?
• Deterministic.
• Fully observed.
• The action space is discrete.
• Access to a perfect simulator (the
game itself), so the effects of any
action are known exactly.
• Each episode/game is relatively short.
• Evaluation is clear, fast and allows a
lot of trial-and-error experience.
• Huge datasets of human play.
• NOT physiology
• NOT disease course
• NOT drug response
• NOT surgery
• So…. What might succumb…? Which
medical ‘game’ fits the criteria….?
• The game of reimbursement
• Has funded billions of $$ of EHR
• Of ontology wrangling companies etc
13. DEPARTMENT OF
Biomedical Informatics
Bias and fairness: AI applications in medicine
• GAN to maximize
reimbursement
• GAN to
minimize/maximize care
utilization
• Opaque biases in drug or
procedure
recommendations.
•At-will review/exploration
of doctor’s
decisions/advice
•Maximize patient’s utilities
in clinical decision-making
over all achievable
treatment plans
•Netflix for medicine
14. DEPARTMENT OF
Biomedical Informatics
AI -> $$$ <- Platform/Path/Payors
Electronic
Health
Record
Data
(labs, meds, images…)
Summary & Decisions
(Rx, Dx, Prognosis…)
$
Consumers
$Employers
$Government
`
16. DEPARTMENT OF
Biomedical Informatics
But at least we have a solid medical education
system…
16
If a test to detect a disease whose
prevalence is 1/1000 has a false positive
rate of 5%, what is the chance that a
person found to have a positive result
actually has the disease, assuming you
know nothing about the person's
symptoms or signs?
19. DEPARTMENT OF
Biomedical Informatics
• Who completes the loop?
• Who is trusted?
• Who integrates with rest of care?
• Who oversees
• Process automation?
• Who does expert catch?
• What do you have to show to get
$$$
21. DEPARTMENT OF
Biomedical Informatics
Reflections/Conclusion
• How would you like to spend your AI effort chits?
• Doing what doctors are supposed to do at expert level?
• Or by what they cannot do (MD super powers)?
• Most of the interesting part of medicine is not in classification or prediction
• It’s in making the right DECISION for precious human beings faced with suffering/death
• Unless you roll a new one, you have to understand THIS healthcare system
processes NOT just medicine art/science.
• Understanding the “job-to-be-done” and where the $$ coming from as important
as CS/AI +medical expertise.
• Doctors are not necessarily expert in this prioritization.
• Don’t lie, it makes us all look bad.
• YOU can roll a new health care system that is data and knowledge-driven.
22. DEPARTMENT OF
Biomedical Informatics
DATAPOWERED STRATEGIESTO
COUNTERANTIBIOTICRESISTANCE
Harvard Medical School
June 8th 2018
Join us for discussions on ‘omics’, health-services, and
internet data integration for innovation in diagnosis and
treatmentThank you
"On Oxford Steet in Cambridge, Mass. lives a sibyl, a priestess of science. Her devotees take their problems to her as devout ancient Greeks took their insolubles to Delphi. She is no mumbling, anonymous priestess, frothing her mouth with riddles. Her name is Bessie; she is a long, slim, glass-sided machine with 760.000 parts, and the riddles that are put to her and that she unfailingly answers concern such matters as rocket motors, nuclear physics and trigonometric functions". For a computing machine, Bessie is old: she has been steadily at work since 1944. And she is not the brightest of her breed. Compared to her children and grandchildren (one of whom, Harvard's Mark III lives on the floor below in Harvard's Computation Laboratory), she is dim-witted and slow. But Bessie is a progenetrix, a sort of mechanical Eve. By proving what computing machines could do, she started one of the liveliest developments of modern science. Some scientists think that Bessie's descendants will have more effect on Makind than atomic energy. Modern man has become accustomed to machines with superhuman muscles, but machines with superhuman brains are still a little frightening. The men who design them try to deny that they are creating their own intellectual competitors." ‒Time Magazine, January 23, 1950, p54.
Important, What can it avoid? Lots of stupidity. Missed diagnoses..
4
umor necrosis factor-α (TNFα) antagonists are effective for the treatment of inflammatory bowel diseases, demonstrating improvement in patients' quality of life, and reductions in surgeries and hospitalizations.1 However, around 10–30% of patients do not respond to the initial treatment and 23–46% of patients lose response over time. Determining whether the reason for failure is a primary or secondary non-response is paramount to successfully treat these patients. A significant proportion of patients do not respond (primary non-response—PNR) to TNFα antagonists. Distinct mechanisms underlie these two forms of TNFα antagonist treatment fail
Not noise
Not confounding
Signal
Different in China than US
Which game in medicine is a fully human artefact with interesting yet somewhat arbitrary rules?
Opaque biases a lot worse than the google image search problems
Instances on left much more $$$ in nesr term.
Where is the money?
3.3 trillion or $10,348 per person
17.9% of GDP
37% Fed
34% Private
11% Out of pocket
Hospital Care 32%
MD/clinical services 20% ($664.9B in 2016)
Prescrption drugs 10% $328B
Home healthcare 3% $93B
Who will be the beneficiary?
https://nccih.nih.gov/research/results/spotlight/americans-spend-billions $30B/year on alternative health