Third Wave AI in Medicine:
From Test-free Screening of Complex Diseases
to
Understanding Microbiome Self-organization and Zoonotic Emergence
Ishanu Chattopadhyay, PhD
Assistant Professor of Medicine
University of Chicago
ishanu@uchicago.edu
mathematics
computer science
social science
medicine
AI/ML learning theory and applications
Complex systems
Implication of AI in Future of Societay
University of Chicago Medicine
The Laboratory for Zero Knowledge Discovery
collaborators
Alex Leow
Psychiatry UIC
Anna Podolanczuk, Pulmonary Care, Weill Cornell
Gary Hunninghake, Pulmonary C, Harvard
Robert Gibbons, Bio-statistics
Daniel Rubins, Anesthesia and Critical Care
Peter Smith, Pediatrics
Michael Msall Pediatrics
Fernando Martinez, Pulmonary Critical Care, Weill Cornell
James Mastrianni, Neurology
James Evans, sociology
Erika Claud, Pediatrics
Aaron Esser-Kahn Molecular Engineering
David Llewellyn
University of Exeter
Kenneth Rockwood
Dalhousie University
Andrew Limper Mayo Clinic
zed.uchicago.edu
Dr. Shahab Asoodeh
Dr. Yi Huang
Dmytro Onishenko
Victor Rotaru
Jin Li
Ruolin Zhang
David Yang
Dr. Nicholas Sizemore
Drew Vlasnik
Lucas Mantovani
Jaydeep Dhanoa
Jasmine Mithani
Angela Zhang
Warren Mo
Kevin Wu
zed.uchicago.edu
Department of Pediatrics
UChicago
Department of Neurology & The Memory Center
UChicago
Department of Psychiatry
UChicago
Pulmonary Critical Care, Weill Cornell
Department of Anesthesia and Critical Care
UChicago
Center for Health Statistics
UChicago
Pulmonary Critical Care, Harvard Medical School
Department of Psychiatry
UIC
Demon Network, Exeter, Alan Turing Institute, UK
Dalhousie University, Canada
Pritzker School of Molecular ENgineering
Social Science
UChicago
zed.uchicago.edu
D3M (I2O)
PAI (DSO)
PREEMPT (BTO)
YFA (DSO)
NIA
ACT 1
Late or missed diagnosis of serious illnesses
Can we use existing EHR to reliably screen for complex diseases such as pulmonary fibrosis, dementia and rare cancers?
Ai
Electronic Healthcare Record
IPF
ASD
ADRD
Onishchenko, Dmytro, Robert J. Marlowe, Che G. Ngufor, Louis J. Faust, Andrew H. Limper, Gary M. Hunninghake, Fernando J. Martinez, and Ishanu Chattopadhyay. "Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records." Nature Medicine 28, no. 10 (2022): 2107-2116.
Universal screening for complex diseases
ACT II
Can We Model Ecosystems As They Evolve ?
Can we predict future mutations?
Digital Twins for complex systems
Nicholas Sizemore, Kaitlyn Oliphant, Ruolin Zheng, Camilia Martin, Erika Claud and Ishanu Chattopadhyay, A Digital
Twin of the Infant Microbiome to Predict Neurodevelopmental Deficits, Science Advances, 2024, In Press
Can we find generative models for microbiome dynamics?
Chattopadhyay, Ishanu, Kevin Wu, Jin Li, and Aaron Esser-Kahn. "Emergenet: Fast Scalable Pandemic Risk Assessment of Influenza A Strains Circulating In Non-human Hosts." (2023). Under Review in Nature
ACT I
The need for Universal Screening
Takes too long,
not supported by insurance,
"gut feeling" / "wait & see" common
Is AI/ML adding anything of relevance?
"predicting" autism > 3yrs
"diagnosing" fibrosis from lung imaging
"diagnosing" dementia from brain scan
Rapid Universal Point-of-care Screening for ILD/IPF Using Comorbidity Signatures in Electronic Health Records
Flag patients before they (or doctors) suspect
Primary Care
Pulmonologist
?
Zero-burden Co-morbid Risk Score (ZCoR)
Rapid Universal Point-of-care Screening for ILD/IPF Using Comorbidity Signatures in Electronic Health Records
shortness of breath
dry cough
doctor can hear velcro crackles
Common Symptoms
>50 years old
more men than women
IPF
Rare disease
~5 in 10,000
Post-Dx
Survival
~4 years
At least one misdiagnosis
~55%
Two or more misdiagnosis
38%
Initially attributed to age related symptoms:
72%
Cannot always be seen on CXR
Non-specific symptoms
PCP workflow demands
~ 4yrs
current survival ~4yrs
~ 4yrs
current clinical DX
ZCoR screening
Onishchenko, D., Marlowe, R.J., Ngufor, C.G. et al. Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records. Nat Med 28, 2107–2116 (2022). https://doi.org/10.1038/s41591-022-02010-y
n=~3M
AUC~90%
Likelihood ratio ~30
Conventional AI/ML attempts to model the physician
AI in IPF Research
ICD administrative codes
IPF
ILD
target codes appear
Past medical history
No target codes appear
case
control
2yrs
2yrs
prediction
target codes appear
Past medical history
No target codes appear
case
control
2yrs
2yrs
IPF drugs prescribed
Signature of IPF diagnostic sequence
pirfenidone or nintedanib
ICD Codes can be noisy
"cases" are not always true IPF
Truven MarketScan (IBM) Commerical Claims & Encounters Database 2003-2018
>100M patients visible
>7B individual claims
>87K unique diagnostic codes
>7% Medicare data present
2,053,277 patients included in study
Univesity of Chicago Medicam Center 2012-2021
68,658 patients
Random sample from Optumlabs Data Warehouse courtsey Mayo Clinic
861,280 patients
2,983,215 patients
Data: Onishchenko etal. Nat. Medicine 2022
performance tables
Marketscan Out-of-sample Results
specificty~99%
NPV>99.9%
IPF
ILD
performance tables
UCM Out-of-sample Results
specificty~99%
NPV>99.9%
IPF
ILD
Comorbidity Spectra
patient A
patient B
patient C
lesson 1
Beyond "risk factors" to personalized risk patterns
False Positives:
Ethics:
For every 20-30 flags,
1 is positive
minimal
acceptable?
Better outcomes
Collard, Harold R., Alex J. Ward, Stephan Lanes, D. Cortney Hayflinger, Daniel M. Rosenberg, and Elke Hunsche. "Burden of illness in idiopathic pulmonary fibrosis." Journal of medical economics 15, no. 5 (2012): 829-835.
Clinical Trial Cohort Selection
Current screen failure rate ~50-60%
ZCoR boosted screen failure rate ~20%
Longitudinal history is important
lesson 2
Off-the-shelf AI does not suffice
lesson 3
Leveraging Longitudinal Patterns
Specialized HMM models from code sequences
Model control and case cohorts seprately
given a new test case, compute likelihood of sample arising from case models vs control models
sequence likelihood defect
ZeD Lab: Predictive Screening from Comorbidity Footprints
Nature Medicine
JAHA
CELL Reports
Science Adv.
1 in 59
Autism Spectrum Disorder
ASD: Ineffective screening causes delays and incurs costs
Autism Co-morbid Risk (ACoR) Score
Data: Onishchenko etal. Science Advances 2021
Autism Co-morbid Risk (ACoR) Score
MCHAT/F
Head to head comparison with current practice
Data: Onishchenko etal. Science Advances 2021
Joint Operation with MCHAT
CHOP Study allows us to see effectiveness of MCHAT in different sub-populations
Modulate sensitivity/specificity trade-offs
Data: Onishchenko etal. Science Advances 2021
The ZCoR Approch: Rapidly Re-targettable
ZED performance | Competition | |
---|---|---|
Autism | >80% AUC at 2 yrs | "obvious" |
Alzheimer's Disease | ~90% AUC | 60-70% AUC |
Idiopathic Pulmonary Fibrosis | ~90% AUC | NA |
MACE | ~80% AUC | ~70% AUC |
Bipolar Disorder | ~85% AUC | NA |
CKD | ~85% AUC | NA |
Cancers (Prostate, Bladder, Uterus, Skin) | ~75-80% AUC | Low |
Deploy all/many/most of these!
>5 Million in US. >13 Million in next 10 years
Alzheimer's Disease and Related Dimentia
MOCA, Blood Tests
Current Practice:
state of art with EHR:
~67% AUC*
ZCoR: ~87%
Alzheimer's Disease and Related Dimentia
state of art with EHR:
~67% AUC*
ZCoR: ~87%
Preempting ADRD accurately upto a decade in future
Application to Suicide Attempts and Ideation (SISA) , PTSD*:
perhaps surprising connection between mood disorders and physiological comorbidities
Gibbons RD, Kupfer D, Frank E, Moore T, Beiser DG, Boudreaux ED. Development of a Computerized Adaptive Test Suicide Scale-The CAT-SS. J Clin Psychiatry. 2017 Nov/Dec;78(9):1376-1382. doi: 10.4088/JCP.16m10922. PMID: 28493655.
* in press
Application to Malignant Neoplasms*
Melanoma
Melanoma has a high survival rate of over 90% when treated early. But if it progresses to later stages, the survival rate drops significantly. Identifying potentially life-threatening melanomas is crucial.
* in press
Reading (References)
Onishchenko, Dmytro, Yi Huang, James van Horne, Peter J. Smith, Michael E. Msall, and Ishanu Chattopadhyay. “Reduced False Positives in Autism Screening via Digital Biomarkers Inferred from Deep Comorbidity Patterns.” Science Advances 7, no. 41 (October 8, 2021). https://doi.org/10.1126/sciadv.abf0354.
Onishchenko, Dmytro, Daniel S. Rubin, James R. van Horne, R. Parker Ward, and Ishanu Chattopadhyay. “Cardiac Comorbidity Risk Score: Zero‐Burden Machine Learning to Improve Prediction of Postoperative Major Adverse Cardiac Events in Hip and Knee Arthroplasty.” Journal of the American Heart Association 11, no. 15 (August 2, 2022). https://doi.org/10.1161/jaha.121.023745.
Onishchenko, Dmytro, Robert J. Marlowe, Che G. Ngufor, Louis J. Faust, Andrew H. Limper, Gary M. Hunninghake, Fernando J. Martinez, and Ishanu Chattopadhyay. “Screening for Idiopathic Pulmonary Fibrosis Using Comorbidity Signatures in Electronic Health Records.” Nature Medicine 28, no. 10 (September 29, 2022): 2107–16. https://doi.org/10.1038/s41591-022-02010-y.
Huang, Yi, Victor Rotaru, and Ishanu Chattopadhyay. “Sequence Likelihood Divergence for Fast Time Series Comparison.” Knowledge and Information Systems 65, no. 7 (March 16, 2023): 3079–98. https://doi.org/10.1007/s10115-023-01855-0.
Brenner, Lisa A., Lisa M. Betthauser, Molly Penzenik, Anne Germain, Jin Jun Li, Ishanu Chattopadhyay, Ellen Frank, David J. Kupfer, and Robert D. Gibbons. "Development and validation of computerized adaptive assessment tools for the measurement of posttraumatic stress disorder among US military veterans." JAMA Network Open 4, no. 7 (2021): e2115707-e2115707.
ACT II
ishanu chattopadhyay
Digital Twin of the Maturing Human Microbiome
Nicholas Sizemore
Kaitlyn Oliphant
Erika Claud
THE PROBLEM
Can microbial assay from gut actionably
pre-empt developmental markers?
Assuming a 1000 species ecosystem, and 1 successful experiment every day to discern a single two-way relationship, we would need 1,368 years to go through all possibilities. If we look for 3 way interactions, we would need 454,844 years
2019
PREEMPT
27 Million
Can we predict the next pandemic?
Can we predict future mutations? Can we define the "edge of emergence"?
Digital Twins for complex systems
Chattopadhyay, Ishanu, Kevin Wu, Jin Li, and Aaron Esser-Kahn. "Emergenet: Fast Scalable Pandemic Risk Assessment of Influenza A Strains Circulating In Non-human Hosts." (2023). Under Review in Nature
PREEMPT
Q-Net
recursive forest
This is a general method!
Data
\(\downarrow \)
Set of interdependent
predictors
How do we measure "distance" between strains?
E-distance
a biologically informed, adaptive distance between strains
smaller distances imply a quatitatively high probability of spontaneous jump
$$J \textrm{ is the Jensen-Shannon divergence }$$
Sanov's Theorem & Pinsker's Inequality
Theorem
stable strain \(x_{h}\), "well-adapted" \(\Rightarrow Pr(x_h\rightarrow x_h) \approx 1 \)
For "new" strain \(x_{a}\), \( \displaystyle \theta(x_{a},x_{h}) \approx 0 \)
Assume:
Then, we have:
we can tell if new strain will adapt to humans
A Math Solution to a Hard Biological Problem
Influenza Risk Assessment Tool (IRAT) scoring for animal strains
Can we replicate IRAT scores*?
slow (months), quasi-subjective, expensive
*https://www.cdc.gov/flu/pandemic-resources/monitoring/irat-virus-summaries.htm
genomic analysis
receptor binding
animal
transmission
antivirals available
population immunity
human infections
animal
hosts
global prevalence
antigenic novelty
disease severity
Influenza Risk Assessment Tool (IRAT) scoring for animal strains
slow (months), quasi-subjective, expensive
*https://www.cdc.gov/flu/pandemic-resources/monitoring/irat-virus-summaries.htm
24 scores in 14 years
~10,000 strains collected annually
Emergenet: finding emergence risk of animal strains
Emergenet time: 1 second
BioNorad
Lets go back to the Microbiome Problem
<class>_<observation_time>
<actinobacteria>_<30wk>
<clostridia>_<28wk>
construct qnet
Sanov's Theorem & Pinsker's Inequality
Theorem
stable profile \(x_{h}\), "well-adapted" \(\Rightarrow Pr(x_h\rightarrow x_h) \approx 1 \)
For "new" profile \(x_{a}\), \( \displaystyle \theta(x_{a},x_{h}) \approx 0 \)
Assume:
Then, we have:
we can tell if new profile will be stable
A Math Solution to a Hard Biological Problem
current state:
also all "future" values for a sample would be missing
typically sparse, lots of missing data
class_time | abundance level |
---|---|
Actionobacteria_28 | a |
Actionobacteria_29 | - |
Actionobacteria_30 | b |
. . . | |
Clostridia_28 | g |
. . . | - |
Bacilli_28 | d |
. . . | |
Gammaproteobacteria_28 | e |
. . . | - |
Coriobacteriia_28 | w |
missing
Biology-aware Perturbations to "reconstruct" missing data
?
Can i meaningfully perturb abundance values?
Can we fill them in if they are missing?
sample
reconstructed observation
sample
a
unknown
current state
a
"collapse"
sample
a
unknown
current state
a
"collapse"
completely uninformative state
No information available for this sample yet
completely uninformative state
No information available for this sample yet
completely uninformative state
observed
state
Q-net inferred with typical patients
Q-net inferred with patients with neurodevelopmental deficit
completely uninformative state
observed
state
Q-net inferred with typical patients
Q-net inferred with patients with neurodevelopmental deficit
Risk of Time-stamped Microbial Profile to lead to Developmental Deficit
smaller the q-distance,
higher the likelihood of a jump
How different are the typical and deficit models?
Actinobacteria 30
typical
deficit
Bacilli 30
typical
deficit
Bacteroidia 30
typical
deficit
Coriobacteria 32
typical
deficit
Gammaproteobacteria 32
typical
deficit
typical
deficit
All Patients
Feeding Variables added
Ability to "fill in" missing data is equivalent to making trajectory forecasts
Our risk measure is highly predictive and actionable
Which entities are most predictive?
Just add those microbes back?
No transplantation is guaranteed to work reliably
Predicted to reduce
risk reliably
Predicted to reduce
risk reliably
Supplantation MUST be personalized
Supplantation MUST be personalized
Supplantation MUST be personalized
Network Interpretations?
Typical
Deficit
Effect of Clinical Variables
Future
Concretely answer the question: "what is a healthy microbiome?"
Explicit supplantation profiles that are tuned to individual ecosystems
Bioreactor experiments
What other problems can it solve?
Q-Nets
Digital Twins for complex systems
Mental health diagnosis
opinion dynamics
algorithmic lie detector
YFA 2020
Yang, David, James EVans, and Ishanu Chattopadhyay. "‘Its the Economy Stupid’: Predictive Theory of Belief Shift Connecting Economic Stress to Societal Polarization." (under review Nature Human Behavior).
predict worldviews from incomplete data
VeRITaAS
Can A Generative AI Tell if you Are Lying?
Vetting Response Integrity from
cross-Talk in Adversarial
Surveys
Hidden structure of cross-talk between responses to interview items
PTSD diagnostic interview
Q-Net
VeRITAS
High Complexity
Low Surprise
Responses that are reflective of symptoms
structured interview
properties of true responses
1
2
3
Minimum AUC = \(0.95 \pm 0.005\)
Cannot be coached, or memorized
Number of possible responses
Minimum Performance (n=624)
Average Time: 3.5 min
No. of Items: 20
AUC > 0.95
PPV > 0.86
NPV > 0.92
At least 83.3% sensitivity at 94% specificity
Beat the test!
Future
Vision
Transform bio-surveillance
Transform modeling of complex systems
Transform early diagnosis
Democratize AI unleashing its power for social good
Why are ML/AI models complicated, and non-transparent?
individual data points not so much important
Tyco Brahe
(1546-1601)
Johannes Keplar (1571-1630)
Newtonian theory of Universal Gravitation (1684)
raw data
empirical fit
universal law of physics
30,000 experiments
Starting point of modern genetics
Mendel's Laws of Genetics
Johann Gregor Mendel (1822–1884)
Some datasets are large, but simple: easily compressible or representable
Others, are not.
"big data" has irreducible complexity
Hence, "models" must have capacity to accommodate this complexity
Machine Learning and AI allows us to find "theories" which are no longer specifiable as simple equations,
but require
billions of parameters to specify
Medical history
co-morbidities
lifestyle
genetics
environment
Estimate disease risk
Estimate prognosis
Reduce missed and delayed diagnosis
Find prodromal patients for clinical trials
The Age of Data
Autism Spectrum Disorder + AI
Idiopathic Pulmonary Fibrosis + AI
Literature Search: AI + Target Disease
Current AI Applications are limited in practice
Are ML predictions pertaining to clinical diagnoses adding anything of relevance?
Risk
The Key Stumbling Block: Features
How to find good features?
Good features
relevant risk factors
Leverage Vast Patient EHR and Insurance Claims Database(s)
Truven MarketScan (IBM) Commerical Claims & Encounters Database 2003-2018
87M patients visible > 1 year
>7B individual claims
>87K unique diagnostic codes
>7% Medicare data present
Cloud Deployment
[
{
"patient_id": "P000038",
"sex": "F",
"birth_date": "01-01-2006",
"DX_record": [
{"date": "07-31-2006", "code": "Z38.00"},
{"date": "08-07-2006", "code": "P59.9"},
{"date": "08-29-2016", "code": "J01.90"},
{"date": "09-10-2016", "code": "J01.90"},
{"date": "11-14-2016", "code": "J01.91"}
],
"RX_record": [
{"date": "10-29-2011", "code": "rxLDA017"},
{"date": "05-16-2015", "code": "rxIDG004"},
{"date": "08-08-2015", "code": "rxIDG004"},
{"date": "06-04-2016", "code": "rxIDD013"}
],
"PROC_record": [
{"date": "02-05-2007", "code": "90723"},
{"date": "11-05-2007", "code": "J1100"}
]
}
]
{
"predictions": [
{
"error_code": "",
"patient_id": "P000012",
"predicted_risk": 0.005794344620009157,
"probability": 0.8253881317184486
}
],
"target": "TARGET"
}
Data In
Data Out
The Paraknowledge API
curl -X POST -H "Content-Type: application/json" -d '[{"patient_id": "P28109965201", "sex": "M", "age": 89, "fips": "35644", "DX_record": [{"date": "12-16-2011", "code": "R09.02"}, {"date": "12-30-2011", "code": "H04.129"}, {"date": "12-30-2011", "code": "H02.109"}], "RX_record": [], "PROC_record": [{"date": "09-28-2012", "code": "71100"}]}]' "https://us-central1-pkcsaas-01.cloudfunctions.net/zcor_predict?target=IPF&api_key=7eea9f70d79c408f2b69847d911303c"
Current Targets
IPF
ILD
ADRD
CKD
CKD_SEVERE
MELANOMA
CANCER_PANCREAS
CANCER_UTERUS
SISA
Cohort Selection and Risk Analysis Testbed
Cohort Selection and Risk Analysis Testbed
Baseline prevalence of IPF in ILD patients
~25%
ZCoR PPV: 60% @ 50% sensitivity
1310 positive patients from 2183 flags
screen failure:
~70% \(\rightarrow\) 40%
Selection comparison against baseline of 2+ ILD risk factors
baseline prevalence: ~2%
projected screen failure:
~98% baseline \(\rightarrow\) 45%
Patient Journeys for IPF: Tracking increasing Risk Over Time
Upto 4 year "signal" resolution
patient journey
Other Examples
decreases risk
increases risk
Risk decreases sometimes
new codes change trajectory as they are revealed
Delving Deeper into Learning Goals
Early screening of complex diseases by leveraging deep pattern discovery in history of medical encounters
Use AI to transform the landscape of early disease diagnosis, prevention, and treatment strategies for complex medical conditions.
Realize universal primary care low-burden screening for disorders for which potentially no recommended screening tools exist currently
Generalize beyond known “risk factors”, uncover personalized predictors of future risk of serious diseases from subtle comorbidity signatures
Problem: Event-level prediction in social systems,
e.g. predicting crime before it happens
Predictive intelligence for security
Ai
Can we predict complex spatio-temporal stochastic processes?
Rotaru, Victor, Yi Huang, Timmy Li, James Evans, and Ishanu Chattopadhyay. "Event-level prediction of urban crime reveals a signature of enforcement bias in US cities." Nature human behaviour 6, no. 8 (2022): 1056-1068.
Problem: Can AI predict how we think and interact?
Can we predict how opinions evolve?
Digital Twins for complex systems
YFA 2020
Can an AI tell if you are lying?
Can an AI tell how you are going to vote?
Yang, David, James EVans, and Ishanu Chattopadhyay. "‘Its the Economy Stupid’: Predictive Theory of Belief Shift Connecting Economic Stress to Societal Polarization." (2023).
Learning Objectives
What is AI/Machine Learning? What are the key application in the context of medicine? What does it bring to the table in the context of Health Services and Bio-medicine? Are there new questions that we can answer? Does it suffice to draw on off-the-shelf models? What are the new/emerging ideas?
Application of AI in Biomedicine: Why We Need a “Bio”-AI.
Emerging tools for addressing Late and Missed Diagnosis in Primary Care
Why “risk factors” are often not predictive enough, and how to think about more personalized predictors of future risk of serious diseases
Zero-burden EHR Analytics
Diagnostic & Screening for complex disorders
*CoR : * Comorbid Risk Scores
ACoR (Autism)
PCoR (IPF/ILD)
ZCoR (ADRD/AD)
ZCoR-C (cancers with further specialization)
Kolmogorov complexity
Surprise
Naive diagnostic Risk
304 VA participants with physician validated PTSD (malingering possibility not considered?)
310 online participants with no mental health diagnosis asked to intentionally malinger
~5% successfully beat the test
~89% of PTSD positive patients pass the test
Substance Abuse Disorder
malingering
SUD
No SUD
Cook County Data
Estimated malingering rate 0.34
650 BCE
1792
1890
1943
1956
2006
2011
2020
2021
2022
Babylonian astrology for prediction
National Weather Service
John McCarthy coins "Artificial Intelligence."
IBM's Watson wins Jeopardy!
AI begins to outperform in healthcare diagnostics
Old Farmer's Almanac first published
McCulloch Pitts' neural network
Deep learning by Geoffrey Hinton
GPT-3
GPT-4 | Dall-e | AI reaching critical mass