Key points are not available for this paper at this time.
Artificial intelligence (AI) has rapidly expanded across clinical decision-making over the past decade, with applications now reported in diagnostic imaging, risk prediction, early warning systems, treatment planning, triage, workflow optimisation, and operational decision support. This narrative comprehensive review, rather than a systematic literature review or meta-analysis, examines how AI supports clinical decision-making and what validation, implementation, bias, and deployment challenges limit its real-world use. Literature was identified through structured searches of PubMed, Scopus, Web of Science, IEEE Xplore, and ScienceDirect for studies published between January 2015 and December 2024 using predefined AI and clinical decision-support terms. Eligible studies reported clinically relevant AI applications with clearly described datasets, modelling approaches, validation strategies, and evaluation procedures. Editorials, commentaries, letters, short abstracts, duplicates, and studies lacking sufficient methodological detail or patient-level clinical relevance were excluded. This review was not registered in International Prospective Register of Systematic Reviews (PROSPERO), did not follow a formal systematic review protocol, and did not apply a formal risk-of-bias tool such as Prediction Model Risk of Bias Assessment Tool (PROBAST) or Quality Assessment of Diagnostic Accuracy Studies-AI (QUADAS-AI); instead, methodological concerns were appraised narratively. The final narrative synthesis included 49 publications, and findings were organised by clinical domain, data source, AI approach, validation method, performance measures, implementation context, and reported limitations. Because of heterogeneity across clinical settings, datasets, model architectures, outcome definitions, comparator groups, and evaluation metrics, pooled quantitative synthesis or meta-analysis was not performed. Reported performance measures included area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, calibration, and predictive values, but these metrics were not pooled because the included studies used different populations, tasks, thresholds, and validation designs. AI showed potential benefits in imaging-based diagnosis, risk prediction, early warning systems, workflow support, and personalised treatment planning, although comparative evidence against non-AI clinical decision support or traditional statistical models was inconsistent and incompletely reported. Diagnostic AI systems were most consistently supported in imaging-based tasks such as mammography, diabetic retinopathy screening, lung nodule detection, and digital pathology, while early warning models showed promise for sepsis, acute deterioration, and critical care risk prediction. Clinical translation remains limited by retrospective and single-centre study designs, inconsistent external validation, limited prospective testing, variable calibration reporting, weak usability assessment, bias, poor transparency, and dataset shift. These limitations restrict certainty about the proportion of deployed AI systems that produce measurable clinical benefit. AI may enhance clinical decision-making, but sustained clinical benefit requires rigorous prospective evaluation, multi-site external validation, fairness assessment, clinician-centred integration, accountable governance, and continuous post-deployment monitoring.
Patil et al. (Fri,) studied this question.