University of Chicago Medicine

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 (2021): eabf0354.

https://docs.google.com/document/d/15iMXe3PkDfRpP9DeBdmGahQ342Aah9OCDWvaMHf4aAQ/edit?usp=sharing

NIH Proposal 1R21MH135316-01

Autism Spectrum Disorder + AI

Literature Search: AI + Target Disease

ML To Enable Universal Screening

• Often the problem is missed or late diagnoses in the primary care workflow
• Universal screening for many diseases are non-existent

Takes too long, not supported by insurance, objective assessments might be difficult

Even when such tools exist they are not administered universally

MCHAT/F (ASD)

Must do pattern discovery

Discover factors that modulate risk, beyond what is already known

Must account for the possibility of non-causal spurious associations

1 in 54

ASD: Ineffective screening causes delays and incurs costs

Autistic children experience higher co-morbidities

Can we exploit these patterns to predict diagnosis?

Common Knowledge: Comorbidties  Exist

ACoR:

Autism Co-morbid Risk Score

D. Onishchenko, Y. Huang, J. van Horne, P. J. Smith, M. M. Msall, I. Chattopadhyay, Reduced false positives in autism screening via digital biomarkers inferred from deep comorbidity patterns. Sci. Adv. 7, eabf0354 (2021).

Autism Co-morbid Risk (ACoR) Score

Autism Co-morbid Risk (ACoR) Score

Automatically infer Importance of different comorbidity categories

17 categories chosen:

immune | infections | endocrine | ...

Retrospective Validation on 2 Dtabases

ACoR Prospective Trial

Patient Flow

Aim 1. Validate reduction of false positives compared to M-CHAT/F

Aim 2. Evaluate dependency between ACoR and M-CHAT/F

Aim 3. Evaluate performance in diverse populations

Aim 4. Characterize heterogeneity of ASD from comorbidity patterns

University of Chicago Women's Board

ACoR Prospective Trial

Patient Flow

1. Pediatric clinic (led by Dr. Mitchell) administers M-CHAT/F to children with 16-26 months,procures consent
2. The PI’s team will compute individual ACoR (steps 1 and 2).
3. If flagged by M-CHAT/F as high risk, or if the ACoR score indicates high risk and M-CHAT/F is borderline, the patients will be scheduled for ADOS-2 evaluation overseen by Dr. Smith & Dr. Msall (Step 3).
4. Evaluation scores will be analyzed by PI and his team.

Michael Msall        Peter Smith         James Mitchell

Conventional AI/ML often attempts to model the physician

Universality

 Universal Screening at the point of care

@ zero additional burden

• Improve patient experience
• Decrease cost to heathcare systems
• Improve physician productivity

Impact

Retrospective Results

Autism Co-morbid Risk (ACoR) Score

Head to head comparison with current practice

Autism Co-morbid Risk (ACoR) Score: Geospatial effect

works well everywhere

Pattern Discovery amidst Heterogeneity

Patterns are discernible in the decisions, but not easily reducible to few human-understandable intuitions

Co-morbidity Spectra

Specific codes that increase the odds of true positive vs true negative

Joint Operation with MCHAT

CHOP Study allows us to see effectiveness of MCHAT in different sub-populations

Modulate sensitivity/specificity trade-offs

ACoR: Variation with Age

can track risk increase over time

Older children are easier to diagnose

Performance on Diverse Populations

Largely insensitive to race/ethnicity

How?

Theoretical Grounding & Publications

• Incorporate longitudinal signals
• Discover unknown factors/patterns that module future risk

Longitudinal patterns matter

The Secret Sauce: Inferring Probabilistic Machines from Data

Immune female control

Immune female case

The Secret Sauce: Inferring Probabilistic Machines from Data

Endocrine female control

Endocrine female case

The Secret Sauce: Inferring Probabilistic Machines from Data

Cardiovascular female control

Cardiovascular female case

Secret Sauce: Leverging Temporal 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

Obvious Features Not Highly Predictive

Sanity checks

Using how often a child is sick is NOT sufficient

Uncertainty in coding does not seem to radically impact results

Good disambiguation of ASD from ID and other disorders