MedAI_for_MDT

The AI will use scattered documentation to prepare background material for the oncology team: for each piece of information, it will provide a verified citation, highlight any discrepancies and identify any missing data.

What MedAI for MDT does. It reads the patient’s entire medical records (OCR-scanned PDFs and exports from NIS) and prepares:

facts (diagnosis, staging, histology, treatment, imaging, laboratory results, current status) – each with a verbatim citation that the code verifies within the document text; without a verified citation, the fact will not be displayed,
a brief summary and a completed ‘Patient Consultation’ form (FNO),
notes: discrepancies between documents (fulvestrant i.v. × i.m., three different entries for the everolimus dose), changes over time (meningioma growth from 10×6 to 12×9 mm), missing data according to the committee’s checklist (histology, dated MR scans, ECOG); if a question for the MDT is missing, mark this and suggest the most likely question as a clearly marked AI proposal,
patient timeline on a single screen: diagnosis → surgery → treatment lines → response on the timeline,
PDF report from fields selected in the application, draft request for further investigations or treatment, and HL7 FHIR R4 export (Bundle for InterSystems IRIS for Health). How it works. Batch processing in advance: LLM extraction → citation verification → merging of identical facts from multiple documents → rules and conflict checking → JSON. The application (ASP.NET Blazor) then displays the case immediately; the LLM is only required for queries during the MDT and for the request proposal. Committee types are configurable profiles (neuro-oncology NOCO, general oncology KOC) – new committee = new JSON, new data source = new loading module. Results based on hackathon data (20 patients). 972 facts, 99.6% with verified citations; all 10 known issues in the data identified; manual annotation of 2 patients: 57/57 facts correct, errors only in the data (96.5% of data correct). Costs approx. 0.12 USD per patient. Next steps. Integration with NIS/LIS/PACS via FHIR, a local model directly within the hospital (architecture is ready), validation with FNO doctors using real-world cases.