2. DEPARTMENT OF
Biomedical Informatics
Artificial Intelligence in Medical Diagnosis
PETER SZOLOVITS, Ph.D.; RAMESH S. PATIL, Ph.D.; and
WILLIAM B. SCHWARTZ, M.D.; ANNALS OF INTERNAL
MEDICINE Vol.108; No.1, pages 80-87. January 1988.
Medicine is an information
and knowledge processing
discipline
9. DEPARTMENT OF
Biomedical Informatics
THE BREAKTHROUGH (2012)
• Pivotal event occurred
in an image
recognition contest
which brought
together 3 critical
ingredients for the
first time:
• Massive amounts of
labeled images
• Training with GPUs
• Methodological
innovations that
enabled training
deeper networks
while minimizing
overfitting
16. DEPARTMENT OF
Biomedical Informatics
Buggy whip is making a space next to it for
pure medical image critiquing
• Not pure critique role may be
enhanced:
• retinal images (might increase
surgical time)
• Mammography and other
screening
• Competition and co-opting for
off-shored image reads
• Pathology
• Dermatology
• Radiology
• Dysmorphology
20. 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?
26. DEPARTMENT OF
Biomedical Informatics
Which path?
Electronic
Health
Record
Data
(labs, meds, images…)
Summary & Decisions
(Rx, Dx, Prognosis…)
Who are likely
winners in for this
path?
27. DEPARTMENT OF
Biomedical Informatics
Which path?
Electronic
Health
Record
Data
(labs, meds, images…)
Summary & Decisions
(Rx, Dx, Prognosis…)
Imaging companies?
AI companies?
ED MD? Oncologist?
What are the levers?
28. 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.
29. 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
31. 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),
32. DEPARTMENT OF
Biomedical Informatics
Big $$$$ questions for next 5 years
• Will AI be primary tool in billing information warfare ?
• Game over for medical imaging?
• How does work and $$$ get redistributed
• Deliver better outcomes?
• Who: Penetrate or bypass existing healthcare process automation?
• AHC vs Google/Apple/Facebook vs
• Will the point of maximum, $$$ leverage be outside or inside the providers’
offices/hospital
• Does the high-touch, shamanistic part of medicine get re-valued,
de-valued or replaced?
33. DEPARTMENT OF
Biomedical Informatics
Thank you
Special Thanks to
Andrew Beam, PhD
Nathan Palmer, PhD
Kun-Hsing Yu, MD, PhD
Bryce Allen, PhD
Ken Mandl, MD, MPH
Bill Geary
Editor's Notes
The opportunity has been clear for ½ century. What I am going to talk about about why the opportunity is not immiment but is here and it’s time to think how it changes medicine and who pays/
Though deal numbers are aaturating, deal size is not: Larger investments coming in health care spending grew 5.8 percent in 2015, reaching $3.2 trillion or $9,990 per person. As a share of the nation's Gross Domestic Product, health spendingaccounted for 17.8 percent.Dec 6, 2016
Though deal numbers are saturating, deal size is not: Larger investments coming in health care spending grew 5.8 percent in 2015, reaching $3.2 trillion or $9,990 per person. As a share of the nation's Gross Domestic Product, health spendingaccounted for 17.8 percent.Dec 6, 2016. Deals are between 0.1 and 0.3% of total spend. For a knowledge-driven business this may be low.
From image applications to identifying modifiable risks and modifying treatment decisions to instrumenting the 99.9% of our lives that most of us spend outside healthcare institutions… no part of medicine and health is untouched.
Quick discloure I am on the SAB of MedAware and HealthReveal.
There is huge and disruptive value in AI in medicine BUT
Is this the way this value is currently captured.
Capturing billions in value today? Really? Let’s look closer
It’s not obvious how the Potential Value is monetized and by whom.
$17-29B annual economic costs in errors ---- even if the errors are reduced, how does that help providers if NOT at risk?
and note that Waste and Fraud is in column 1 and Improved Billing Effiiency is column 2. Both sides will not come out ahead.
Furthermore, there are issues of tense confusion: future vs present that lead to some fundamental and perhaps premature disappointments.
So let’s hone in on where the real excitement in in AI
After fundamental insights in the 1950s on artifical neurons and limitations articulated by Minsky and Papert in the 1960’s to 1970’s
solid methodological progress occurred in background but many of us were shocked by a breakthrough even in 2012 in an image recognition contest.
The insight I want to be sure you take away is that this success was profoundly an matter of successful engineering.
Apart from the methodology: engineering success #2 is massive data sets with high quality labels.
Third engineering successin in computing infrastructuture: To feed violent fantasies of pimply teenage nerds graphic board for computers became more powerful than the central processing units. So rather than 4 or 8 cores: 1000’s
11 teraflops 3584 Cuda cores
Eart Simulator Super computer e site officially opened on March 11, 2002. The project cost 60 billion yen. 35.86TFLPS (4 of these)
Before ever playing a real game DeepStack went through an intensive training period involving deep learning (a type of machine learning that uses algorithms to model higher-level concepts) in which it played millions of randomly generated poker scenarios against itself and calculated how beneficial each was. The answers allowed DeepStack’s neural networks (complex networks of computations that can “learn” over time) to develop general poker intuition that it could apply even in situations it had never encountered before. Then, DeepStack, which runs on a gaming laptop, played actual online poker games against 11 human players. (Each player completed 3,000 matches over a four-week perio
So how about the human contact part of medicine…
So my mother was hospitalized twice in 3 months 2 years ago to get fluid out of her body with IV diuretics because of heart failure. I was determined to prevent this costly (to her) event from recurring and despite many studies showing failure of digital health interventions in heart failure…
. Ex vivo linear suturing under deformations. The experiment consisted of closing a longitudinal cut along pig intestine, whereas the tissue was deformed by pulling on stay sutures. Five samples were tested per technique (OPEN, LAP, RAS, and STAR). (A) Suture spacing. Central mark is the median; box edges are the 25th and 75th percentiles, error bars are the range excluding outliers, and red dots are outliers. The whiskers represent the range not including outliers. There is a different N number for each boxplot because each surgeon used a different number of sutures [OPEN (n = 174), LAP (n = 128), RAS (n = 176), and STAR (n = 206)]. These data are presented numerically in table S2, including the SDs. P values determined by ANOVA with post hoc Games-Howell. (B and C) Leak pressures and number of mistakes (repositioned stitches or robot reboot). Data are from individual tissue samples (n = 5) with averages marked by a horizontal line. P values determined by independent samples t test. (D) Completion times separated into knot-tying and suturing, and other time was spent restaging or changing sutures. Data are averages (n = 5). P values determined by independent samples t test.
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