Artificial Intelligence in Oncology

Our Research

Artificial intelligence is transforming how we understand human disease. Our research develops machine learning methods and foundation models that integrate large-scale biomedical data to investigate disease mechanisms, predict health trajectories, and advance biomedical discovery.

We develop AI methods that learn from diverse data modalities, including electronic health records, pathology images, genomics, molecular profiling, spatial biology, and medical imaging. One of the lab’s flagship research programs is Delphi, a foundation model for human disease.

Our research combines methodological advances in artificial intelligence with applications to fundamental and clinically relevant questions. We develop multimodal learning approaches, large language models, and computational methods for integrating heterogeneous biomedical data. Current research includes disease prediction from longitudinal health records, AI-based digital and computational pathology, prediction of molecular tumour characteristics from routine histology, integration of pathology with genomic and imaging data, and AI-assisted approaches for biomedical research and diagnosis.

Working closely with clinical and experimental collaborators, we aim to use advances in artificial intelligence to generate new insights into cancer and other diseases and to contribute to more precise, data-driven biomedical research and diagnostics.

Current Projects

Developing and Refining Delphi

Delphi is a probabilistic generative model of health trajectory that is trained on large-scale electronic health record data and jointly models the risks of more than 1000 human diseases. Bridging across temporal point process models, epidemiology, and survival analysis, we aim to develop Delphi into a research platform that achieves accurate and calibrated disease risk prediction and serves as a discovery engine for human health, revealing how diseases co-occur, cascade, and progress over a lifetime. Active research areas include but are not limited to multimodal extensions – biomarkers, genetics, and imaging data, cross-cohort generalizability and robustness, novel point process architecture, and explainability.