Ishanu Chattopadhyay PRO
ML | Data Science Biomedical Informatics | Social Science | Assistant Professor
Predicting Adverse Effects for Antifibrotic Therapy
in
Pulmonary Fibrosis
Using Large Digital Twins
Dmytro Onishchenko
Institute of Biomedical Informatics
Ishanu Chattopadhyay, PhD
Assistant Professor of Internal Medicine
Institute of Biomedical Informatics
&
Computer Science
University of Kentucky
ishanu_ch@uky.edu
Non-specific Symptoms
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%
- >50 years old
- more men than women
shortness of breath
dry cough
doctor can hear velcro crackles
Cannot always be seen on CXR
PCP workflow demands
Poor Diagnosis
- Late/missed diagnosis
- No universal point-of-care screening
Pulmonary Fibrosis
Solving the Diagnosis Problem with AI
~ 4yrs
current survival ~4yrs
~ 4yrs
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
2. Conventional AI attempts to model the physician
1. Use of AI in point-of-care diagnosis is limited
- Use under-utilized diagnostic modalities
- Discover co-morbidity patterns
- Go beyond know risk factors
- Personalized risk patterns
Off-the-shelf AI
State of the art screening approaches are inadequate
Onishchenko, Marlowe, Ngufor, Faust, Limper, Hunninghake, Martinez, and Chattopadhyay. "Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records." Nature Medicine 28, no. 10 (2022): 2107-2116.
AI: Under The "Hood"
| Drugs | Adverse Effects |
|---|---|
| Pirfenidone | - Nausea, vomiting, diarrhea - Photosensitivity, rash - Fatigue, dizziness - Liver enzyme elevations - Anorexia, decreased appetite - Headache |
| Nintedanib | - Diarrhea, nausea, abdominal pain - Liver enzyme elevations - Decreased appetite, weight loss - Increased bleeding risk - Hypertension - Headache, fatigue |
Poor Prognosis
- No disease modifying drugs
- Anti-fibrotics only slow-down disease
- Poorly tolerated
Adverse Effects in PF
Heterogeneous presentation of adverse effects
\}
\}
Not a classification problem
Too little data to directly train on discontinuation cohort
Digital Twin of Health Trajectory
G72_61 G72_60 G24_62 G81_62
patient1 9 x x ''
patient2 x 2 '' ''Encoding all medical encounters
G72.9 occurs at age 61 for patient 1
We do not know if G81 occurred at age 62
38000 ICD-age features
20K patients with 50% having high IPF risk
Risk-balancing with ZCoR
Train a digital twin of diagnostic trajectories
individual diagnostic trajectories encoded
Large Health Model
Digital Twin of Health Trajectory
G72_61 G72_60 G24_62 G81_62
patient1 9 . . ''
patient2 . 2 '' ''Encoding all medical encounters
individual diagnostic trajectories encoded
A04_53
C50_56
Z85_56_2
950402 parameters
Average tree depth: 3.28
Risk-balanced training data
Probabilistic tracking of future diagnoses across the human disease spectrum
Large Health Model
Pr(x \in \mathsf{CASES} \vert \{ r_i(x), i=1,\cdots, m\}) \geqq \left ( \prod_{i=1}^m PPV_i(t=r_i(x)) \right )^{\frac{1}{m}} \rho^{1-\frac{1}{m}}
Use LHM to calculate risk with respect to all ICD codes:
r_i(x)
For patient \(x\) for disease \(i\):
Define the minimum risk from a set of predictors
\left ( \prod_{i=1}^m PPV_i(t=r_i(x)) \right )^{\frac{1}{m}} \rho^{1-\frac{1}{m}}
prevalence
Positive predictive value
Risk Definition from multiple predictors
Predictive Performance Results
Predictive Performance Results
Restricting predictors to commonly expected adverse effects of antifibrotic therapy
Predictive Performance Results
Optimizing the risk predictor set using in-sample performance
Code definition of Pulmonary Fibrosis
CONSORT diagram
Conclusions
- Substantial improvement of adverse-effect prediction
- Discovery of "new" risk factors
- Demonstrate realistic simulation of health trajectories leading to new capability
Future
Further development of "Large Health Model" framework as a generalizable generative tool for digital twin inference in clinical problems
antifib_DT_CCTS
By Ishanu Chattopadhyay
