Predictive Medicine
Overview
Predictive medicine applies machine learning to answer the clinical question: what will happen to this patient? Rather than diagnosing the current state, predictive models estimate the probability and timeline of future events — disease progression, treatment response, or transition to a more severe stage.
Clinical applications
Diabetic Retinopathy Progression
Diabetic retinopathy is the leading cause of blindness in working-age adults globally. My research builds end-to-end deep learning pipelines that process sequential retinal fundus images and OCT scans to predict whether a patient will progress to proliferative DR or develop diabetic macular oedema within a given follow-up window.
Age-Related Macular Degeneration (AMD)
AMD progression from early/intermediate to late neovascular AMD is a critical decision point for preventive treatment with anti-VEGF agents. Models trained on multi-modal imaging data from the MARIO dataset can identify high-risk patients up to 12 months before conversion.
MARIO Challenge
I co-organised the MARIO AMD Progression Monitoring Challenge at MICCAI 2024, providing the community with a standardised benchmark and evaluation protocol for AMD progression prediction from longitudinal OCT imaging.
Methodology
The pipeline typically involves:
- Pre-training on large unlabelled imaging datasets (self-supervised)
- Fine-tuning on labelled longitudinal cohorts with survival or classification objectives
- Calibration to produce reliable probability estimates for clinical decision support