Research portfolio

Heart transplant patients undergo multiple endomyocardial biopsies each year for rejection surveillance, even though most biopsies are negative — creating unnecessary procedural risk, cost, and patient burden. This project integrates biopsy histopathology representations with ECG foundation-model embeddings in a multimodal deep learning framework for rejection risk prediction after cardiac transplantation. EMB slides are encoded with an attention-based multiple-instance learning model, ECGs are represented as 768-dimensional foundation-model embeddings, and a gated fusion network combines both modalities into a continuous patient-level risk score.

• Fusion model improved C-index vs. EMB-only model: 0.66 vs. 0.61 on a held-out test set (N=134)
• Significant Kaplan-Meier risk separation with fusion HR=2.97 (95% CI: 1.47-6.00); likelihood-ratio test p=0.0017
• Published as a 2026 Journal of Heart and Lung Transplantation abstract

Heart transplant recipients currently require 15–20 endomyocardial biopsies per year for rejection surveillance — an invasive procedure carrying procedural risk and significant cost. This project applies ECG foundation model embeddings to longitudinal 12-lead ECG series, learning subtle temporal patterns that precede rejection episodes. The approach achieves 15.1% biopsy reduction at 95.9% sensitivity, compared to 6.9% with the standard clinical decision rule, and identifies a 3× elevated 90-day rejection hazard ratio in flagged high-risk patients.

• 15.1% biopsy reduction vs. 6.9% clinical baseline at 95.9% sensitivity
• 3× elevated 90-day rejection hazard ratio in high-risk group
• Accepted for presentation at ISHLT 2026

CKD patients face 5–10× elevated cardiovascular risk, yet standard risk calculators (e.g., AHA-PREVENT) perform poorly in this high risk population group. ARiSe repurposes routine retinal fundus photographs — already captured at optometry visits — to extract vessel geometry and vascular texture features that encode systemic vascular disease burden. A multi-feature risk model built on these retinal biomarkers outperforms the PREVENT calculator (C-Index: 0.70 vs. 0.65) in a prospective cohort of 800 CKD patients (CARE-CKD study). The same retinal feature pipeline was also evaluated for Alzheimer's disease detection, demonstrating cross-disease applicability.

• C-Index 0.70 vs. 0.65 (AHA-PREVENT) validated on 800 CKD patients
• Cross-disease utility: retinal features applied to Alzheimer's detection
• Published in Kidney360 (2025)

Non-alcoholic fatty liver disease affects ~25% of the global population yet is rarely screened for in asymptomatic patients. Millions of low-dose CT scans are acquired annually for coronary calcium scoring — a routine cardiovascular test — while the liver, clearly visible in the same scan, is routinely ignored. DeHFT/LARI is a deep learning and radiomics pipeline that automatically quantifies liver fat from these incidentally acquired scans, enabling opportunistic NAFLD screening at no additional imaging cost or radiation dose. Validated across 700+ patients in multi-site cardiovascular outcomes cohorts; results published in Lancet eBioMedicine (2025) and underpinning an allowed US patent.

• Validated on 700+ patients across multi-site cardiovascular outcomes cohorts
• Published in Lancet eBioMedicine (2025)
• US Patent allowed (US 20240420320 A1)

Standard echocardiography reports reduce a full cardiac cycle to a handful of static measurements, discarding the temporal dynamics — how the heart contracts and relaxes over time — that encode important prognostic information. STAR-Echo is a spatiotemporal biomarker derived by fusing Vision Transformer-based sequence embeddings with handcrafted radiomics features extracted from echo video clips. Trained on CKD patients with longitudinal follow-up, it achieves AUC 0.71 for predicting major adverse cardiovascular events (MACE) and produces interpretable spatial attention maps identifying the cardiac regions driving the prediction.

• AUC 0.71 for MACE prediction in CKD patients
• Vision Transformer + radiomics fusion architecture with interpretable attention maps
• Published at MICCAI 2023 | Pending US Patent (US 20250118435 A1)

As Co-President of the BME Graduate Student Advocacy Board, I represent PhD students in policy and welfare discussions with department leadership at Georgia Tech and Emory. Focus areas include TA compensation structures, health insurance access, and PhD program transparency.

• Organized student-faculty townhalls on PhD training and career development
• Restarted the Mock Quals mentorship program pairing senior and junior PhD students
• Student voice on TA compensation and health insurance policy