Transforming oncology clinical trial matching through multi-agent AI and an oncology-specific knowledge graph: A prospective evaluation in 3,800 patients.

Abstract

1554 Background: Clinical trial enrollment in oncology is often hampered by the manual, time-intensive process of matching patients to trials with highly specific eligibility criteria. Advances in artificial intelligence (AI)—particularly multi-agent large language models (LLMs) and oncology-specific knowledge networks—hold promise for streamlining this workflow and minimizing human labor. This abstract presents a prospective evaluation of an AI platform that automates medical data extraction, leverages an oncology-specific knowledge graph, and provides real-time trial recommendations, demonstrating a significant reduction in staff effort while maintaining high clinical accuracy. Methods: Multi-Agent AI & Oncology Knowledge Graph 1. OncoAgents: Specialized LLMs (data extraction, eligibility, trial matching), collaborating to outperform generic or zero-shot AI. 2. OncoGraph: Domain-specific knowledge graph uniting patient data, molecular profiles, and clinical guidelines for context-aware matching. 3. OncoRecommend: Real-time engine processing new data, trials, and guidelines, delivering rapid, relevant suggestions. 4. OncoSet: Expert-curated dataset (>2,000 patient records, 14,000+ trials, 50+ tumor subtypes) ensuring robust AI performance. Prospective Analysis (Jan–Dec 2024): Cohort: 3,804 patients (ECOG 0–2) with metastatic/progressing malignancies seeking trial options. Data Extraction: 157,367 pages (~86.5M tokens) processed for tumor type, stage, treatment lines, and biomarkers. Trial Matching: Automated application of inclusion/exclusion criteria; oncologists validated AI-generated matches. Efficiency: Manual matching for large cohorts can require thousands of hours; this AI approach condensed it to ~1 hour of expert review. Results: 1. Screening & Identification: 3,804 patients screened; 23,912 trials identified; 17,912 confirmed after expert review. 2. Time-to-Recommendation: Under one week from screening to final recommendations via real-time AI prioritization. 3. Performance Metrics: Sensitivity (Recall): 0.8375, Specificity: 0.8359, Precision: 0.8121, F1 Score: 0.8246. Demonstrates advantages over zero-shot or frontier GPT-based models. 4. GPT Comparison: Extraction Accuracy: 80.29% vs up to 63.15% (GPT-4o); Trial Matching Accuracy: 82.06% vs 47.00% (GPT-4o). Conclusions: This multi-agent AI platform, underpinned by an oncology-specific knowledge graph, significantly boosts efficiency and accuracy in oncological trial matching. By cutting manual workloads from thousands of hours to near-automated speeds, recommendations allow for just-in-time, decentralized and patient-centric trial activation. Ongoing enhancements—such as deeper biomarker integration, expanded knowledge graph coverage, and seamless EHR interoperability—promise further gains in personalized oncology care.