ASCOmind: Is instant ASCO abstract analysis possible with AI agents?

Abstract

7557 Background: The ASCO Annual Meeting receives thousands of abstracts annually on ongoing therapies. Extracting actionable insights from this large volume of data through manual review is time-consuming. To reduce manual workload and accelerate evidence synthesis, we implemented an AI-Agent system to assess the feasibility of deploying AI agents for efficient, large-scale data analysis and insight generation. Methods: GPT4o-based ASCOmind was designed with a robust framework of six autonomous and collaborative AI agents: Pre-processor, Categorizer, MetadataExtractor, Analyzer, Visualizer, and ProtocolMaker to systematically generate and visualize insights from ASCO abstracts. We demonstrate and evaluate ASCOmind by applying it to 2024 multiple myeloma (MM) studies. Using human reviewers as the gold standard, we assessed the quality and efficiency of the system focusing on outcome data accuracy, visualized charts, and workflow recipes documentations. Results: ASCOmind processed abstracts in the plasma cell dyscrasia section, categorizing 60 MM abstracts into 26 clinical trials and 34 as real-world studies. Manual abstraction of 51 predefined data elements required >60 mins/abstract, whereas ASCOmind completed the same task in <5min/article. The ASCOmind not only significantly reduced the processing time but also instantly analyzed and visualized the extracted data within 10 min. For instance, 27 included high-risk populations with cytogenetic abnormalities (n=20), extramedullary disease (n=7), or elderly patients (n=4). Across 51 interventional studies, 33 targeted relapsed/refractory MM (RRMM) and 18 focused on newly diagnosed MM (NDMM). ASCOmind generated a treatment distribution table for RRMM and NDMM (Table 1), with one misclassification corrected by humans-Mezigdomide reclassified from ADC to the correct category. Additionally, granular efficacy/safety outcome values and summarized study findings were also successfully extracted and visualized. Conclusions: Our preliminary analysis of ASCOmind demonstrated high accuracy and efficiency in automating abstract analysis, enabling rapid analysis of trends and outcomes. This feasibility study highlights the scalability of AI systems across all cancer types, supporting decision-making. Treatment distribution in RRMM and NDMM studies. Total (N=51) Therapy Category Examples No. of Studies % RRMM(N=33) BCMA-CAR-T Therapies Cilta-cel, Ide-cel, ARI0002h 8 24.3% BCMA-Bispecific Ab Therapies Teclistamab, Talquetamab, Elranatamab, Linvoseltamab, ABBV-383 16 48.5% ADC Belantamab mafodotin, Elotuzumab, 5 15.2% Cereblon E3 Ligase Modulator Mezigdomide, Iberdomide, 2 6.0% Others OriCAR017, Venetoclax 2 6.0% NDMM(N=18) Triplet/Quadruplet SOC VRd, Isa-VRd 8 44.5% Transplantation ASCT, Tandem Transplantation 6 33.3% ADC Belantamab mafodotin, 2 11.1% BCMA-Directed Therapies Cilta-cel + Lenalidomide, Teclistamab 2 11.1%