Research
Active Projects
Adaptive Modeling for Personalized Lab Reference Intervals
Redefining "normal" with patient-specific baselines
Routine blood tests are compared against population-wide "normal ranges" that miss individual variation. I build hybrid models that blend population priors with personal history, dynamically updating reference intervals as more data arrives — catching early disease signals that standard ranges miss.
4.9M+ Patients
20+ yrs Longitudinal Data
3 Modeling Approaches
Replacing Race in Pulmonary Function Test Reference Equations
ARC & ARCPFT: anatomy over demographics
Lung function tests have long used race-based equations with no biological basis. I developed ARC, a framework that identifies anatomical proxies (sitting height, body morphology) to explain group-level lung function differences — and ARCPFT, a new reference equation that improves accuracy and equity simultaneously.
~160K Participants
24% Error Reduction (Black patients)
2 Cohorts (NHANES + UK Biobank)
Shortcut Learning in Vision-Language Models for Medical Imaging
Probing GPT-4V and Gemini Pro for clinical robustness
As VLMs enter clinical workflows, I systematically benchmark their reliability. Minor prompt rephrasing bypasses safety constraints and elicits confident diagnoses. Both models show shortcut learning and reduced accuracy on underrepresented subgroups — highlighting critical gaps in prompt-robustness and fairness.
3 Imaging Domains
2 Foundation Models