Group · Predictive
Machine learning for biomedicine, built for deployment in clinics that cannot absorb a black box. A model we cannot explain to a clinician is a model we do not consider finished.
Risk prediction and patient stratification from multi-modal electronic health records — early deterioration, readmission, and progression — validated against the shift between the data a model was trained on and the population it will meet.
Convolutional networks and vision transformers applied to radiographic, histopathological and neurological imaging for computer-aided diagnosis, with careful attention to how little labelled data is available locally.
Feature attribution, saliency and counterfactual explanation, treated as a requirement rather than an appendix. Where a model cannot be made interpretable, we report that as a limitation on its use.
Ensemble learning and feature selection to isolate diagnostic and prognostic signatures from omics data, with the selection stability reported alongside the performance.
To be listed