Projects

Built an interpretable ML classifier for Brugada syndrome on the Brugada-HUCA 12-lead ECG dataset, extracting 111 features per record across all 12 leads. Trained XGBoost and Random Forest models under repeated stratified 5-fold cross-validation, achieving 0.90–0.91 AUROC and 76% sensitivity at 90% specificity. Used SHAP attribution to explain predictions and tie misclassifications to a clinically recognized concealed ECG phenotype.

Built a CNN-based arrhythmia classifier trained on MIT-BIH and evaluated for cross-dataset generalization on PTB-XL, using AAMI beat taxonomy and SCP-ECG label harmonization.

Built a reproducible pipeline classifying brain MRI slices into glioma, meningioma, pituitary, or no-tumor categories, with a segmentation scaffold for voxel-level tumor mapping on BraTS. Documented intended use, training data, and limitations in a model card, scoping the project as research-only with no PHI per regulatory and ethics review.

Quantified how much signal quality a passive single-lead thigh ECG loses relative to a clinical 12-lead reference, and how much a preprocessing cascade recovers. Designed and ran four experiments across MIT-BIH and tOLIet datasets, correcting the original study design after discovering the data lacked beat-level labels and sample synchronization.

Built a computational triage pipeline for ClinVar variants of uncertain significance in orphan disease genes lacking a ClinGen expert curation panel. Benchmarked REVEL, CADD, and AlphaMissense against existing ClinVar classifications using ClinGen SVI-calibrated thresholds to generate a prioritized curation list for expert review.

Reanalyzed a genome-scale CRISPRi Perturb-seq screen covering roughly 22 million primary human CD4+ T cells across four donors, profiled at rest and after TCR/CD28 stimulation, to nominate novel and druggable targets.

A network reanalysis of the Gordon/Krogan SARS-CoV-2 protein interaction map (Nature, 2020), covering 26 viral proteins and 332 human host targets.