Nooruldeen
Taffar

A 12-lead ECG is read by a clinician in two movements: locate each lead on the page, then read the waveforms across them. My work automates both, and returns standardised, machine-readable diagnoses rather than free text.

GenECG-Diagnostic pairs a YOLO object detector that crops the twelve lead panels from a raw 3×4 grid image with a Vision Transformer (google/vit-base-patch16-224) fine-tuned under BCEWithLogitsLoss to emit multiple SNOMED-CT codes per ECG. Built end to end — dataset construction, augmentation, YOLO label generation, training for both stages, and a full image-to-diagnosis inference pass.

The wider thesis fuses that visual modality with tabular EHR and demographic data at tensor level, recalibrated toward sensitivity because a missed cardiac event costs more than a false alarm. Evaluation follows TRIPOD+AI reporting and PROBAST+AI risk-of-bias scoring; I am building an instrument for the latter, since appraising medical foundation models is not yet well served.

The clinical case is concrete: cardiovascular diagnostic referrals cost NHS England roughly £369m a year at £300–1,500 each. A decision-support system running alongside conventional workflow, not in place of it, is where I intend to take this — through a PhD in medical informatics from approximately 2027.