COMFORT — Computational Models FOR patienT stratification in urologic cancers

Prostate and kidney cancer are diagnosed using fragmented data: imaging (MRI, CT), free-text radiology and pathology reports, lab values, and patient history — assessed largely in isolation rather than combined. COMFORT builds AI models that fuse these modalities to mimic how physicians actually reason across sources, producing predictions on cancer type, progression, and prognosis. The consortium is running a prospective clinical study across multiple European sites: 685 patients with suspected prostate cancer and 315 patients with suspected kidney cancer, with models trained on clinical data from participating hospitals combined with public sources (medical literature, textbooks).
Within this structure, the DATEXIS team at BHT works on two work packages. In WP3, the team builds NLP models that extract structured clinical information — diagnoses, findings, staging — from unstructured electronic health records and radiology/pathology reports, and integrates these models into a clinical application usable at the point of care. In WP5, the team develops methods for model trustworthiness: explainability of predictions for clinical users, detection and mitigation of bias across patient subgroups, and privacy-preserving processing of patient data — a requirement both for clinical acceptance and for the project’s planned regulatory pathway toward CE marking as a medical device under the EU Medical Device Regulation (MDR).
This builds directly on DATEXIS’s prior clinical NLP work — including Medbert.de, developed with COMFORT coordinator Keno Bressem (Charité) — and continues in the team’s 2026 publication on evaluation robustness for clinical LLMs, co-authored with Prof. Dr. Alexander Löser and Conor Fallon.