Research

My work follows a model through its life — trained, evaluated, deployed — with one discipline underneath all of it: measurement. Each stage asks its own version of the same question: does this number mean what we think it means?

Most of the code behind this work is proprietary to Truveta. The papers, abstracts, and write-ups linked below are its public record.

01 · Train

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

Does a generated patient journey contain the diagnoses and biomarkers that real ones do?

On Truveta's Intelligence ML team I train oncology-specific and general-purpose clinical foundation models on longitudinal electronic health records. Because the training data are messy, the events that matter most — specific diagnoses, key biomarkers — are the ones a plain generative model gets wrong. So I am adding reinforcement learning that pulls generated patient journeys toward real ones.

Clean inputs are part of training, not a preliminary: the same models impute missing units of measure and derive health-state transitions from the data rather than by assumption. I also build ML and LLM risk models for cancer and heart failure.

02 · Evaluate

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

Does the score mean what we think it means?

Benchmarks and model-as-judge systems are instruments that rarely report their psychometric properties. I build evaluations that establish validity and reliability before a metric is used to make a claim. For our patient-journey foundation model, that means a multi-level evaluation framework, from checks of the model itself to the quality of what it generates.

Its benchmarks widen in stages, from one cancer to all cancers to diseases across the general population. Simulations test whether generated journeys follow clinically plausible progression.

In heart failure, EHR-based estimation of KCCQ scores turns on whether a model-derived score means what the patient-reported instrument means. Earlier, at Harvard, I tested LLM graders against human scores for high-stakes essay scoring, the reliability question behind any model-graded evaluation, and helped develop rcttext, an R package for analyzing text outcomes in randomized trials.

03 · Deploy

Tooling decides whether a good model ever reaches anyone

Will the people who need the signal see it in time, and in context?

A model that only runs in a notebook has not been deployed. I build the layer that carries one into use: MCP tools for clinical NLU that pull structured disease status and patient trajectories out of messy real-world text, and agent frameworks — multi-agent and deep-agent architectures, LangChain pipelines — that interpret what the models return.

Around them sit automated prompt optimization, interactive dashboards, and end-to-end pipeline monitoring, so the people making decisions can interrogate a prediction rather than just receive it. Most of the distance between a promising result and a used one is infrastructure.

Related: CV·Write-ups

04 · Foundations

Classical measurement theory and modern machine learning improve each other

What are we actually measuring?

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