Research

My research develops statistical methods and open-source software for valid inference when AI/ML predictions are used in place of observed data, and applies deep learning tools to clinical problems. I also have a body of applied work in public health across maternal health, child welfare, and infectious disease.

Inference on predicted data

As AI/ML models increasingly replace traditional data collection, researchers face a new challenge: how do you draw valid statistical conclusions when your outcome variable was predicted by an algorithm rather than directly observed? Naive use of predicted data produces biased estimates and invalid confidence intervals.

I work on methods that correct for this, and co-develop the ipd R package, which implements multiple approaches (PostPI, PPI, PPI++, PSPA) in a unified framework. Our current extensions target survival/censored outcomes (the VIPS project) and longitudinal data analyzed with generalized estimating equations.

Key papers:

Deep learning for clinical screening

I am developing a multimodal convolutional neural network that combines chest X-ray embeddings and 12-lead electrocardiogram waveforms to screen for pulmonary embolism using the MIMIC-IV critical care database. The model achieves 93.9% sensitivity using only objective, routinely collected data, comparable to existing clinical decision rules that rely on subjective physician judgment.

This project is built using scorcher, an R package I co-develop that makes deep learning accessible to clinical researchers without specialized programming backgrounds. Scorcher is supported by a Sloan Precision Oncology Technology Dissemination Award.

Collaborators: Dr. Jeffrey T. Leek (Fred Hutch), Dr. Stephen Salerno (Fred Hutch), Dr. Barbara Lam (Fred Hutch/UW), Dr. Leo Anthony Celi (MIT/Harvard)

Biomedical software evaluation

As first author, I led an assessment of impact metrics for NCI-funded cancer research software through the ITCR program, finding that traditional citation counts inadequately capture software usage and that infrastructure like documentation quality and developer accessibility are associated with higher adoption.

Key paper: Best practices to evaluate the impact of biomedical research software – Bioinformatics, 2024

Applied public health

I have published on a range of public health topics using nationally representative survey data from Bangladesh, including maternal health (antenatal care, cesarean sections, menstrual hygiene), child welfare (stunting, child abuse), intimate partner violence, and COVID-19 (vaccine prioritization, pandemic preparedness, underreporting). See my full publication list for details.