Feasibility assessment of a global, anonymized, human-in-the-loop AI decision-support platform for oncology second opinions following a live demo at an international symposium.

Abstract

e13672 Background: There are widespread disparities in access to rapid, expert-level second opinions worldwide. We assessed the feasibility and early adoption of an anonymized, human-in-the-loop AI decision-support platform (PrecisCa.AI) that analyzes cases in 2-3 minutes and offers optional expert faculty over-read on request. Methods: After a real-time demonstration during international breast cancer tumor boards (December 2025), we collected provider-submitted, de-identified cases. Feasibility measures included successful case processing and measured time to response. Secondary measures included clinician uptake, location of submitted cases, tumor of origin, referral sources, and in-app feedback. AI consults are undergoing expert review and scoring from 1-5 (5 being the highest) on 6 attributes: clarity, completeness, menu of options, recency, reasoning, and relevance. An in-app survey asked if the response was helpful. Results: Across the observation window, from 12/10/2025-1/23/2026, 256 cases were submitted by 89 unique active providers (mean 2.9 cases/provider) after 125 new sign-ups (100% registration completion). Processing success was 100% (0/256 failed), with mean turnaround time of 149 seconds. Submissions originated from North and South America, Europe, Middle East, Africa, and Asia. Breast cancer cases (58%) were most common followed by thoracic 13%, hematologic 9%, GI 8%, GU 5%, gynecologic 3%, head & neck 2%, and skin 1%. Referrals were most commonly from a colleague (45%) followed by referrals from a conference 32%, email 7%, social media 6%, Google 3%, and other 7%. In-app survey response rate was 16.8% (43/256) with mean score 3.7/5. 33% of survey responses contained free-text comments. Multiple adjudicator large language models were available and used (most commonly recent Gemini and ChatGPT). Conclusions: An anonymized, clinician-submitted, human-in-the-loop AI platform achieved feasible global use shortly after a public demonstration, with 100% processing success and short mean turnaround time across diverse tumor types and referral channels. Optional expert faculty over-read was available, and AI consults will now be expert-rated (1-5) on clarity, completeness, menu of options, recency, reasoning, and relevance. Detailed analyses of expert ratings and comparative model performance will be presented.