A Retrospective Analysis of Clinical Outcomes Following Human Deviations from Algorithmic Recommendations

Abstract

Background: Artificial intelligence (AI)-based decision-support systems are increasingly used in perioperative care to support risk assessment, resource allocation, and surgical planning. However, these systems may not capture qualitative clinical information or “soft data” recognised through experienced clinical judgement. The consequences of senior surgeons deliberately overriding algorithmic recommendations remain insufficiently understood, particularly regarding patient safety and operational efficiency. Aim: To evaluate clinical and operational outcomes when surgeons override AI recommendations and assess the "Disagreement Dividend" to inform healthcare governance. Study Design: Retrospective, multi-centre, propensity-matched cohort study. Place and Duration of Study: Tertiary healthcare centres with mature AI decision-support systems (>1-year post-deployment; IRB approved October 2025). Methodology: Of 1,582 screened cases, 142 senior-surgeon override cases (Cohort B) were propensity-matched (1:2) with 284 AI-compliant control cases (Cohort A) based on patient age, ASA score, procedure complexity, and surgeon experience. Key endpoints included 30-day complications, length of stay (LoS), operating room (OR) time deviations, and qualitative analysis of the rationale for overrides. Results: Override cases showed significantly lower 30-day complication rates (8.5% vs. 14.8%; P = 0.032; aOR = 0.51, P = 0.024). In AI-compliant cases, OR time was underestimated by an average of 18.4 minutes, whereas surgeon timing estimates were more accurate (+3.1 minutes; P < 0.001), and LoS was lower than predicted (-1.2 days; P < 0.001). Overrides were predominantly driven by uncodified "soft data", such as tissue friability (42.2%). Conclusion: Senior clinical intuition accounts for qualitative "soft data" not captured by AI. Expert overrides were associated with improved safety and operational efficiency, supporting human-in-the-loop governance rather than rigid algorithmic compliance.