Clinical AI Beyond Development: A Scoping Review of Deployment-Related Robustness, Algorithmovigilance, and Lifecycle Oversight

Résumé fourni par la source

BACKGROUND/OBJECTIVES: Clinical artificial intelligence (AI) is increasingly moving from proof-of-concept development into clinical evaluation, regulatory review, and routine care. This scoping review aimed to map and synthesise empirical evidence on clinical AI evaluation after model development, focusing on deployment-related robustness, post-development monitoring, and lifecycle oversight in practice. METHODS: We conducted a scoping review in accordance with Joanna Briggs Institute guidance and reported findings using PRISMA-ScR. MEDLINE, Embase, Scopus, and Web of Science Core Collection were searched with no lower date restriction within each database's available indexed coverage and with a common upper search date of 28 February 2026. Searches were supplemented by backward and forward citation tracking. Grey literature, preprint servers, and regulatory databases were not systematically searched because eligibility was restricted to full-text, peer-reviewed empirical studies and empirically grounded implementation or monitoring reports. Findings were synthesised using descriptive evidence mapping and inductive thematic synthesis. RESULTS: Eighteen studies or empirically grounded reports were included. Evidence was organised into five strata: direct live or post-deployment monitoring studies; near-live bridge studies generating prospective outputs without guiding care; methodological monitoring and maintenance studies; deployment-relevant robustness and predeployment safety studies; and governance, implementation, readiness, and human-factors studies. Three themes emerged: trustworthiness after development was conditional and context-dependent; algorithmovigilance extended beyond aggregate performance tracking to include operational, workflow, fairness, contextual, and user-feedback signals; monitoring was more actionable when linked to corrective pathways, governance structures, and institutional readiness. Sociotechnical failures included automation-bias signals, workflow burden, reasoning-conclusion misalignment, and workflow-fit problems. CONCLUSIONS: Post-development clinical AI evaluation remains a layered and emerging field rather than a mature monitoring literature. Direct live evidence is limited, concentrated in high-income settings, and weighted towards radiology. The findings should be interpreted as synthesis-informed rather than as empirically validated standards for lifecycle oversight.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.