Abstract PS3-04-10: Development and Validation of a Living Decision Support Tool (Living-DST) in Oncology Using Agentic AI-Augmented Systematic Literature Review (SLR)

Abstract

Abstract Background: Oncologists face growing difficulty in staying current with rapidly evolving data from congresses, journals, guidelines, and regulatory updates. Creating a well-organized, annotated clinical trial (CT) data library in an easily digestible format is time- and labor-intensive. To address this, we developed and validated a living-DST for breast cancer (BC), powered by an agentic AI system that supports daily human conducted / AI augmented SLR integrated with guidelines, regulatory approvals and ongoing CT results. Methods: An agentic AI system using GPT-4.1 & o3 (OpenAI), and Claude Sonnet-4 (Anthropic) was designed to emulate expert-led Cochrane-compliant SLR processes. The system follows an annotation manual, decomposes tasks into subtasks, and self-debugs and validates. The annotation manual was constructed by human scientists manually annotating 29,236 clinical trial abstracts across BC, lung, and prostate cancer from 2019-current. Each study was annotated with 4 review variables (population, intervention/comparator, reported outcomes, study design) and 32 extraction variables: i.e. TNM staging, histology, biomarkers, risk factors, treatment line, intervention, study design, size, follow-up, outcomes (overall and progression/disease-free survival, response, QoL), subgroup data, and toxicity. Structured data were integrated with national guideline-based treatment pathways, including new evidence beyond guidelines, forming a real-time, evidence-linked DST. Accuracy was assessed vs. 1,997 human annotations as validation set. DST evidence precision was evaluated against the AI chatbots: ChatGPT, Perplexity, and Consensus using 6 criteria for 8 cases on treatment options in BC with different treatment path, subpopulation, and biomarker variables. Accuracy for each criterion was defined as the percentage of questions evaluated as positive over all eight questions. Results: Accuracy across review variables ranged from 95.1-97.2%. For extraction variables, accuracy exceeded 90% for all variables (91.5%-99.1%) with 50% of variables above 95%. Agentic AI annotated 1,997 publications in 7.25 hours vs. an estimated 727.4 hours required by human experts—yielding 99% of time savings. When compared to other AI tools, our DST outperformed by providing more complete and accurate evidence, retrieving recently published, potentially practice-changing pivotal trials (23% of all evidence), reporting all available clinical evidence, and linking to original citations and FDA labels where available (Table). Conclusion: By integrating human-conducted, AI-augmented daily SLR integrated with guidelines and regulatory approvals, our Living-DST delivers real-time, clinically actionable decision support. This system offers oncologists precise, patient-specific insights and may improve cancer treatment outcomes. Citation Format: L. Schwartzberg, H. S. Rugo, A. Forsythe, D. Flora, S. Glück, S. Grieve, R. Campden, R. Liu, J. M. Rege, P. A. Kaufman. Development and Validation of a Living Decision Support Tool (Living-DST) in Oncology Using Agentic AI-Augmented Systematic Literature Review (SLR) [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-10.

Ähnliche Arbeiten