Introduction
Rare Gynecological Tumors (RGTs) represent a significant challenge in oncology due to their low incidence, which precludes large-scale randomized controlled trials. Consequently, clinical guidelines are often sparse, and patient outcomes remain poor. While Molecular Tumor Boards (MTBs) aim to provide precision medicine by matching patients to therapies based on genetic biomarkers, they are hindered by a "data bottleneck." Critical patient information and relevant clinical evidence are frequently trapped in unstructured formats, such as narrative clinical notes and a vast, ever-expanding body of medical literature.
Methodology: A Two-Tier LLM Workflow
To address this curation challenge, researchers developed a proof-of-concept system utilizing Large Language Models (LLMs) to construct "digital twins"—virtual representations of individual patients that integrate clinical and molecular data. The workflow employs a strategic two-tier architecture:
- Privacy-Preserving Local Extraction: A local instance of Gemma-2-27b-it was deployed within the hospital's secure environment. This model extracts sensitive data points—including diagnosis, biomarker status, and treatment history—from institutional Electronic Health Records (EHRs) without compromising patient privacy.
- Large-Scale Literature Synthesis: Gemini-1.5-Pro was utilized to process public medical literature. Its extensive context window allows for the efficient synthesis of hundreds of scientific publications and clinical trial records (e.g., from PubMed and ClinicalTrials.gov) to identify "patients like mine" and relevant off-label therapeutic options.
Results: Efficiency and Accuracy
The LLM-enabled pipeline demonstrated a transformative impact on data curation speed. For institutional EHR data, the automated approach was 6.3 times faster than manual extraction (8.69 minutes vs. 55 minutes per file). Literature curation saw a 2.8-fold increase in speed.
Beyond speed, the models maintained high levels of accuracy. The cloud-based model achieved a precision of 0.97 on literature extraction, while the local model reached 0.85 on complex, unstructured EHR data.
Clinical Impact: Individualized Treatment Targets
The utility of the digital twin approach was validated through a cohort of 21 cases (7 institutional and 14 from literature). A notable success involved a 77-year-old patient with metastatic uterine carcinosarcoma (UCS). By shifting from an organ-centric to a biology-centric view, the system identified high PD-L1 expression and suggested pembrolizumab—an immunotherapy typically used in other cancer types.
The digital twin pipeline also identified several other potential biomarkers and corresponding treatments for this patient, providing a comprehensive map for individualized care.
Following the recommendation for pembrolizumab, the patient achieved a sustained partial response lasting over 30 months, significantly outperforming the typical 1.8-month progression-free survival (PFS) associated with standard third-line treatments for UCS.
Discussion and Future Directions
While promising, the study identified areas for improvement, particularly in temporal reasoning. Extracting the exact timing of treatments and calculating PFS from fragmented notes remains challenging for current LLMs. Additionally, the reliance on English-centric models highlights a need for high-quality, language-specific medical LLMs to capture nuances in local medical dialects.
Future work will focus on integrating this pipeline into larger national networks, such as the German Network for Personalized Medicine (DNPM), to validate the predictive power of digital twins across a broader spectrum of rare cancers. By acting as a "curation accelerator," LLMs can empower clinicians to move beyond traditional protocols toward truly personalized, biomarker-driven oncology.
Summary updated