Research Focus
Four interconnected areas guiding my exploration of machine learning systems engineered specifically for healthcare, with primary emphasis on Domain Shift and Out-of-Distribution Generalization.
Domain Shift & OOD Generalization in Medicine
Cross-scanner variations, demographic shifts, and clinical reliability
When a medical diagnostic model trained on scans from Hospital A (e.g., Siemens MRI scanners with specific slice thicknesses) is deployed at Hospital B (e.g., GE or Philips scanners with different sensor calibration and patient demographic distributions), performance frequently degrades catastrophically without warning. My core research investigates domain shift in clinical machine learning — designing invariant representation learning, test-time adaptation (TTA), and calibrated uncertainty quantification so algorithms recognize when input data is out-of-distribution before emitting an erroneous clinical diagnosis.
Deep learning & neural networks from scratch
Mathematical autograd, backpropagation, and loss dynamics
Deep understanding begins when you construct tensors, matrix calculus, backward passes, and activation landscapes by hand. Reconstructing fundamental architectures without high-level library abstractions guarantees intuition when debugging gradient vanishing, unstable activations, or numerical underflows in complex medical models.
Computer vision for diagnostics
Screening, imaging analysis, and lesion localization
Medical imaging (X-rays, CT slices, histopathology) presents challenges absent in standard natural image benchmarks: severe class imbalance, extreme resolution requirements, and a zero-tolerance margin for missed anomalies. Focusing on CNN architectures, Vision Transformers (ViT), and attention heatmaps (Grad-CAM) to make diagnostic predictions interpretable to clinicians.
Applied ML pipelines at scale
Large, messy real-world datasets with verifiable reproducibility
Hospital and research data is rarely clean: missing modalities, temporal irregularities, inconsistent patient encodings, and massive scale. Developing resilient ETL and tensor ingestion pipelines capable of processing tens of millions of rows while guarding against subtle data leakage between train, validation, and test splits.
Domain Shift & OOD Generalization in Medicine
Core Premise & Approach
When a medical diagnostic model trained on scans from Hospital A (e.g., Siemens MRI scanners with specific slice thicknesses) is deployed at Hospital B (e.g., GE or Philips scanners with different sensor calibration and patient demographic distributions), performance frequently degrades catastrophically without warning. My core research investigates domain shift in clinical machine learning — designing invariant representation learning, test-time adaptation (TTA), and calibrated uncertainty quantification so algorithms recognize when input data is out-of-distribution before emitting an erroneous clinical diagnosis.