Research

Four threads. Two are about building systems that predict something useful from messy clinical data; two are about knowing whether to believe them. They are not separable problems, and treating them as one job is the whole approach.

Foundation models trained on patient journeys can predict what happens next — if the modeling respects the data

On Truveta's AI & Workflows team I train oncology foundation models on longitudinal electronic health records, and build the LLM and machine learning systems that turn them into risk predictions. TrajOnco, a training-free multi-agent framework, does temporal reasoning over sequential clinical events and reaches AUROCs of 0.64–0.80 across 15 cancer types with no task-specific training; a zero-shot chain-of-agents variant reaches 0.871 for one-year lung cancer risk.

Alongside the risk models sit clinical NLU pipelines that pull structured disease status and patient trajectories out of messy real-world text. And the same patient-journey models do less glamorous but equally necessary work: imputing missing units of measure at 98% accuracy across 438 UCUM units, and deriving health-state transition matrices empirically rather than by assumption. High-dimensional temporal EHR data is not clean, and a model that pretends otherwise will be confidently wrong.

An evaluation is a measurement instrument, and should be validated like one

Benchmarks and model-as-judge systems are instruments that almost never report their psychometric properties. I build the frameworks that establish validity and reliability before a metric is used to make a claim: evaluator calibration, benchmark quality, bias detection, and uncertainty quantification — including clinical validity evaluation for the oncology foundation models above, and EHR-based estimation of KCCQ scores in heart failure, where the whole question is whether a model-derived score means what the patient-reported instrument means.

Earlier: the rcttext R package for bringing NLP into randomized controlled trials at Harvard, and a comparative study of large language models against tree-based machine learning for high-stakes essay scoring — which is the same question wearing different clothes.

Tooling decides whether a good model ever reaches anyone

A model that only runs in a notebook has not been deployed. I build the layer that carries one into use: agent frameworks — multi-agent and deep-agent architectures, LangChain pipelines — alongside automated prompt optimization, Model Context Protocol servers, interactive data visualization, and end-to-end pipeline monitoring, so that the people making decisions can interrogate a prediction rather than just receive it.

This is engineering, not methodology, and I think it is underrated. Most of the distance between a promising result and a used one is infrastructure.

Classical measurement theory and modern machine learning improve each other

My doctoral work was on factor analysis and prediction algorithms, and I keep returning to the seam between them: gradient boosting to improve factor-retention decisions under measurement error, penalized estimation (Lasso, adaptive Lasso, and the minimax concave penalty) for exploratory measurement modeling, and network analysis for quantifying item process data.

Latent variable models were built to handle noisy, partial, indirect observation of something you cannot measure directly. That is also a fair description of an electronic health record.

Related: AERA 2024·IMPS 2023·NCME 2021