Publications

Cardiac-toxicity substudy of the 11-286 thoracic radiation cohort relating pulmonary-artery radiation dose to post-treatment pulmonary hypertension. I built the RPDR extraction: echocardiogram and right-heart-catheterization report counts before and after radiotherapy, regex extraction of tricuspid regurgitation severity, TAPSE and PASP from echo reports, and main pulmonary-artery diameter from chest CT reports, delivered to the co-authors for deidentified analysis.

Relates radiation dose to the thymus and lymphocyte health to survival in two lung-cancer chemoradiation cohorts (Harvard-CRT, 422 patients with complete dose data, and the chemoradiation-plus-durvalumab cohort). I extracted and validated the clinical, dosimetric, genomic and lymphocyte data from the 11-286 REDCap database and RPDR laboratory files: performance status, heart and lung V5, PTV volume, LAD dose metrics, tumor genotyping, pre- and post-treatment lymphocyte counts, and durvalumab cycle dates.

FAHR-Face is a foundation model trained on >40 million facial images and fine-tuned for two tasks: biological age estimation (FAHR-FaceAge) and survival risk prediction (FAHR-FaceSurvival). FAHR-FaceAge achieved a mean absolute error of 5.1 years on public datasets and outperformed benchmark models across the full human lifespan. FAHR-FaceSurvival demonstrated robust mortality prediction, with the highest-risk quartile showing more than triple the mortality of the lowest (adjusted HR 3.22; P<0.001). The combination of both models improved prognostic accuracy. Deployed at faceage.bwh.harvard.edu.