Research

Agentic AI Systems for Data Processing & Clinical Risk Prediction

Real-world EHR data are messy: missingness, human error, outliers, and workflow effects can dominate signal and undermine clinical prediction. My research develops agentic AI systems that automatically diagnose data quality issues, execute reproducible preprocessing and feature engineering, and generate clinically grounded representations for modeling, while recording provenance and decision rationales for auditability and interpretability. These pipelines also include agents that select modeling strategies, configure hyperparameter tuning and validation, and adapt workflows to the dataset and clinical endpoint under explicit constraints (e.g., leakage control and robustness checks). I apply these methods across structured data, unstructured clinical text (e.g., notes and discharge summaries), and imaging for perioperative risk prediction (including elective spine surgery outcomes) and for predicting problematic opioid use / opioid use disorder.

Clinical NLP for Substance Use Phenotyping

Substance use disorders are often under-coded in structured EHR fields, with key evidence embedded in clinical narratives. My research develops clinical NLP and phenotyping frameworks that extract and organize substance-use–related signals from notes and discharge summaries while preserving clinical context (e.g., negation, temporality, uncertainty, and attribution). A major focus is problematic opioid use/opioid use disorder, where I build context-aware annotation and modeling pipelines to distinguish clinically meaningful patterns such as active use vs historical use, treatment and recovery markers, and severity-related indicators. These phenotypes are designed to be reproducible and evaluation-ready, supporting downstream prediction, cohort discovery, and longitudinal risk profiling.

Knowledge Graphs & Human-AI Collaboration

Clinical AI workflows often suffer from weak “traceability”: it can be difficult to reconstruct how raw EHR data became model inputs, why specific preprocessing or modeling decisions were made, and what evidence supports a given prediction. My research develops knowledge-graph–based frameworks that encode relationships among clinical concepts, data quality operations, feature definitions, model configurations, and validation evidence. By organizing this information in a structured, queryable form, these graphs support transparent iteration and human–AI collaboration, enabling clinicians and data scientists to discover new clinical insights, compare alternative pipeline choices, and perform feature selection for downstream predictive modeling. I also build graph-informed, retrieval-augmented interfaces that surface provenance and supporting context alongside model outputs to improve interpretability, auditability, and practical decision support.

Modeling Clinical Heterogeneity

Patients with the same diagnosis or procedure can have markedly different risk trajectories, driven by differences in comorbidity burden, prior history, medications, and care pathways. When these sources of variability are ignored, models may perform well on average while failing in clinically important subgroups or time periods. My research focuses on heterogeneity-aware prediction: identifying clinically meaningful subgroup structure, quantifying outcome variability, and evaluating whether model performance is stable across strata defined by baseline clinical profiles and data availability. I use clustering and representation learning to discover risk phenotypes, and I emphasize calibrated risk estimates and error analysis to understand where a model generalizes and where it breaks. This work supports more reliable risk stratification and more actionable clinical decision support in perioperative and longitudinal settings.

AutoML for Genotype–Phenotype Association Analysis

Many genotype–phenotype association studies emphasize additive effects, but biologically meaningful variation can be non-additive, arising from alternative inheritance models and epistatic interactions among loci. My research develops automated machine learning systems for association analysis that move beyond a single encoding of genetic effects by systematically evaluating multiple inheritance models (e.g., additive, dominant, recessive, and other non-additive parameterizations) and searching for higher-order feature interactions. These pipelines prioritize scalable discovery while maintaining rigorous validation, reproducibility, and interpretable summaries of learned genetic signals. The goal is to improve power to detect complex genetic architectures and to provide more mechanistic, interaction-aware representations of genetic contribution to phenotypic variability. Our newest tool, StarBASE-GP, is built to ingest large-scale genomic variant data and assess non-additive effects, including epistatic interactions.