Comparing artificial intelligence and multidisciplinary tumor board decision making in real-world cancer care: a...

Multidisciplinary tumor boards (MDTs) integrate expertise, enhance diagnostic accuracy, improve adherence to evidence-based guidelines, and facilitate individualized treatment planning. Recent advances in artificial intelligence (AI) help streamline these processes, though prospective real-world evaluations remain limited.

We conducted a prospective, non-interventional, blinded concordance study at a tertiary cancer center. Consecutive cases discussed at institutional MDTs between January 2025 and June 2025 were screened for eligibility. Cases with comprehensive clinical information and documented MDT decisions were included. Anonymized vignettes were input into ChatGPT® using a standardized template. A blinded expert reviewer assessed concordance using a predefined three-point scale. The primary outcome was mean concordance score (MCS) for primary clinical query. Secondary outcomes were domain-specific concordance and reviewer-perceived clinical acceptability of AI decisions.

A total of 106 cases (median age 53 years) were analyzed, spanning 21 tumor sites, with the most common being breast (17%), ovary (10.4%), and esophagus (8.5%). Disease stages at MDT discussion were early (20.8%), locally advanced (49.1%), de novo metastatic (8.5%), and recurrence (21.7%). The most frequent primary query was treatment intent and multidisciplinary care (66.1%). MCS for the primary clinical query was 1.42, corresponding to ∼71% of the maximum possible concordance on a 0-2 scale. The highest concordance was noted for treatment intent (84.6% full concordance), radiotherapy (73.4%), and surgical decisions (71.8%), and the lowest for diagnostic (51.6%) and systemic therapy (56%). AI decisions were perceived clinically acceptable for real-world application in 78.3% of cases, comparable to MDT decisions (75.5%).

In a prospective real-world setting, moderate-high concordance was observed for AI and MDT decisions across multiple domains. These findings support further evaluation of AI as a decision-support tool within multidisciplinary oncology care.