IMPACT-OC

IMPACT-OC builds an interpretable, standards-first decision-support stack not only for patients with ovarian cancer but also for indeterminate adnexal lesions and women with unknown hereditary risk for developing ovarian cancer. Calibrated modality-specific agents (imaging, pathology, genomics, clinical text) and a super-agent fuse evidence with guideline logic, site policies and explicit uncertainty to produce auditable recommendations. A longitudinal digital twin forecasts trajectories and short what-if scenarios. Outputs surface via a clinician dashboard and report agent. Privacy-preserving scale-up is central: data stay on site; federated learning exchanges encrypted updates; safety-gated synthetic datasets support benchmarking and stress tests. All assets interoperate through DICOM/DICOMweb for imaging and HL7 FHIR/OMOP for structured data, with DCAT-AP metadata, provenance and model cards. Multilingual NLP and adapted foundation models are packaged as lightweight services. Validation is strictly non-interventional (TRL-4), using retrospective and time-shifted feeds for internal–external testing, calibration, uncertainty quality and fairness audits across centres, scanners, languages and subgroups—without influencing care. Three pilots anchor evaluation: expertise-aware triage between primary surgery and neoadjuvant chemotherapy in advanced HGSOC; ultrasound-based preselection for HRD/BRCA testing; and calibrated multimodal risk for adnexal-mass management. Data access from four European centres enables a three-period plan: foundations (ethics, harmonisation, baseline agents), integration (digital twin v1, dashboards, shadow-mode replay) and final validation (explainability, robustness, fairness). Outputs—open interfaces, interoperability profiles and CE-readiness materials—support adoption by hospitals and SMEs, aligned with the EU AI Act, Medical Device Regulation, EUCAIM and EHDS.