1ST PLACE, ORIGINAL RESEARCH TRACK: CUPNavigator: A multi-model CDSS for tumor origin prediction, diagnostic planning, and therapy recommendations in cancer of unknown primary

Cancer of Unknown Primary (CUP) remains one of the deadliest global health challenges of the 21st century. According to the MD Anderson Cancer Center, CUP cases constitute 2% to 5% of all cancer cases. CUP is a metastatic condition where the tumor origin remains unknown despite extensive imaging, biopsies, and molecular testing, leading to unnecessary procedures, increased costs, diagnostic delays, toxic therapies, and poor outcomes. Simultaneously, diagnostic and treatment guidelines are continuously evolving, burdening oncologists to stay up-to-date and interpret these regimens. To address both tumor origin ambiguity and guideline complexity, CUPNavigator, an AI Clinical Decision Support System (CDSS) was created, integrating (1) a tumor origin classifier trained on somatic mutation profiles from the MSK-MetTropism dataset, (2) a diagnostic agent using Retrieval Augmented Generation (RAG) and NICE guidelines, and (3) a therapy agent grounded in ASCO therapy guidelines. The classifier achieved 90% accuracy across 27 cancer types, outperforming state-of-the-art CUP classifiers. The RAG agents generate site-specific diagnostic steps and therapeutic recommendations that completely minimize hallucinations and maximize guideline adherence in metastatic breast cancer. The entire pipeline executes under 10 minutes, reducing clinical turnaround times by 8 weeks. CUPNavigator provides a longer survival period of 12. 5 months by enabling earlier diagnosis and tumor site-specific therapy and is projected to save 24000 by eliminating toxic side effects. Once clinically validated and deployed, CUPNavigator serves as a valuable tool for navigating CUP challenges and assisting oncologists in saving many lives in clinical workflows.