Governance and Performance Dashboards for Managing Healthcare Analytics and Artificial Intelligence Implementation Projects in Hospital Systems: A Critical Review

Abstract

Hospital systems are adopting predictive analytics and artificial intelligence at a pace that has outstripped the development of the organisational machinery required to supervise them. Two bodies of scholarship have grown in parallel to address this problem. The first proposes governance frameworks that specify principles, lifecycle stages, decision points and oversight bodies for artificial intelligence adoption. The second, considerably older, examines performance dashboards as instruments for making clinical and operational performance visible and actionable within hospitals. These literatures rarely engage with one another, and the practical consequence is that governance frameworks specify what ought to be overseen without establishing how oversight is instrumented, while the dashboard literature describes display and feedback mechanisms without addressing the distinctive properties of algorithmic systems that degrade silently, distribute performance unevenly across patient subgroups and depend on outcome labels that arrive long after predictions are made. This critical review examines the intersection of these fields, drawing on peer-reviewed literature identified through structured searching of bibliographic metadata registries and biomedical indexes to 1 June 2026. Five arguments are developed. Governance frameworks remain concentrated at the level of principles, with oversight mechanisms the least frequently specified component. The indicator base for algorithmic performance is dominated by discrimination metrics whose selection is seldom justified and which are poorly aligned with the decisions that governance bodies actually make. Evidence that dashboards improve care is suggestive rather than conclusive, derives largely from non-algorithmic quality improvement contexts, and depends heavily on how feedback is designed and embedded in organisational routines. Post-deployment monitoring of deployed models is constrained by label latency, by the absence of agreed thresholds for action and by the cost of ground-truth ascertainment. Generative and ambient applications are stretching an oversight model constructed for discrete risk-scoring tools. Priorities for research include comparative evaluation of oversight architectures, development of decision-relevant indicator sets, and prospective study of dashboards as governance interventions rather than as reporting artefacts.