Research Projects

Domain-specific LLM system for decision support in molecular tumour boards (MTBs)

Problem: Precision oncology aims to deliver personalized treatment recommendations based on genomic biomarkers. However, the translation of genomic data into individualized treatment plans is hindered by the fragmentation of medical information across numerous unstructured databases. Delays in diagnostic workflows and the clinical interpretation of biomarker profiles often prevent the timely translation of molecular insights into actionable treatment decisions, see Lancet Oncology Editorial von August 2024(link is external). Previous studies(link is external) have demonstrated that general-purpose large language models (LLMs) are not well-suited for precision oncology therapy recommendations due to their lack of domain specificity.
Our Solution: The MEREDITH (Medical Evidence Retrieval and Data Integration for Tailored Healthcare) LLM system employs Retrieval-Augmented Generation (RAG) to generate therapy recommendations by integrating data from clinical trials, oncological guidelines, therapeutic availability, and trial databases. This system mirrors the approach of experts in Molecular Tumor Boards (MTBs) by interpreting molecular data within a comprehensive medical context.
Status: MEREDITH is capable of assisting precision oncologists in selecting molecularly targeted therapies and identifying suitable clinical trials. Our proof-of-concept study, conducted using fictitious patient data, demonstrated a high level of agreement (94.7%) with human MTB experts.
Details: Publication (JCO Precision Oncology)
A multicenter, investigator-initiated trial will validate our results for real-world use in gynecological precision oncology.

Digital Twins in Gynecological Precision Oncology

Problem: Over 50% of all gynecological tumors are classified as rare (incidence of <6 per 100,000 women). The rarity and heterogeneity of these diseases result in a lack of comprehensive clinical guidelines, hindering the quality of care for affected patients. The conventional approach of classifying metastatic tumors by their site of origin often leads to the exclusion of patients with rare tumors from emerging treatment options. Uterine carcinosarcomas, for example, are included in the S3 guidelines for endometrial carcinoma(link is external), with treatment options limited to first-line therapy with carboplatin/paclitaxel. Despite this, median survival has remained poor for decades, typically less than two years. There is an urgent need for novel, clinically evaluated therapeutic strategies.
Our Solution: Through our work, we aim to catalyze a paradigm shift towards individualized, biomarker-driven treatment strategies. This approach facilitates more precise patient stratification, providing the foundation for personalized therapy—particularly in scenarios where large-scale clinical trials are not feasible. To this end, we have developed and piloted a Digital Twin System: utilizing LLM-supported data extraction, structuring, and analysis of both institutional and literature-based data, we have successfully analyzed a cohort of over 400,000 patients and assessed a range of biomarker-based therapeutic options.
Details: Publication (npj digital medicine)
Status: Launched in May 2025, the GO-TWIN (Gynecological Oncology –​​​​​​​ Targeting Women's Individual Needs) project builds on this proof-of-concept study to improve therapy and quality of life for patients with ovarian cancer. It will use multimodal, large language model-supported analysis of hospital data to develop personalized treatment approaches. Led by Principal Investigator Dr. Lammert, the project is a collaboration between TUM University Hospital, Technical University of Munich, and the University of Regensburg, supported by a €3 million grant from the Bavarian State Ministry of Health, Care, and Prevention.