A context-constrained oncology agent for automated, guideline-concordant treatment decision support in precision oncology.

Abstract

1620 Background: Precision oncology requires biologically plausible, clinically executable decisions aligned with regulatory approvals. General-purpose Large Language Models (LLMs) often produce recommendations rejected by molecular tumor boards due to semantic drift, inappropriate agent extrapolation, and ignored clinical constraints, raising significant safety concerns for point-of-care adoption. Methods: GenOnco is a constrained oncology agent integrating NCCN, ASCO, and ESMO clinical practice guidelines; FDA/EMA regulatory approvals; NEJM seminal articles and high-impact publications; curated clinicogenomic resources (e.g., Molecular Oncology Almanac); comprehensive clinical trials knowledge (ClinicalTrials.gov, landmark phase III results); and practice-changing trial evidence. The system utilizes a dual-layer architecture: a reasoning LLM paired with a deterministic, non-LLM clinical decision layer that enforces line-of-therapy requirements and actionability filters independent of generative output. We evaluated GenOnco on 102 multi-institutional oncologist-derived real-world solid tumor queries, independently scored by two blinded oncologists (κ = 0.87). Performance was benchmarked against published datasets: MEREDITH (200 cases) and Google DeepMind AMIE (100 cases). Primary endpoints were guideline concordance and clinical executability, analyzed via McNemar's test (N = 162). Results: GenOnco achieved 95.1% accuracy (95% CI: 91.2%–98.4%) on synthetic benchmarks and 93.1% (95% CI: 89.1%–96.5%) on real-world queries, eliminating 100% of expert-rejected outputs from comparators. GenOnco significantly outperformed GPT-4o (48.1%), Claude 3.7 (49.4%), Gemini 2.5 (50%), ASCO Guidelines Assistant (67.3%), OpenEvidence, MEREDITH, and Google DeepMind AMIE (all P < .001; N = 162), with gains of 15-47 percentage points. Conclusions: Constrained reasoning prioritizing executable pathways outperforms unconstrained LLMs for precision oncology support. Limitations include solid tumor focus and query diversity; diverse population validation planned. GenOnco demonstrates potential for point-of-care integration.