Scoping review maps machine learning tools reshaping clinical decision support

Machine learning models are increasingly embedded in clinical decision support systems, promising sharper diagnoses and faster treatment choices, but a new scoping review concludes that the technology’s real-world impact hinges less on algorithmic accuracy than on trust, transparency and health system readiness. The review, published in the Bulletin of the National Research Centre, mapped evidence on machine learning within clinical decision support systems, examining their performance, clinical utility, adoption, explainability and implementation challenges in a single synthesis.

The researchers followed PRISMA-ScR guidelines and searched PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore and CINAHL from database inception to 30 March 2026. Guided by the Population–Concept–Context framework, they included studies applying machine learning within clinical decision support in any healthcare setting worldwide. From 3,993 records initially identified, 581 duplicates were removed, leaving 3,412 for title and abstract screening; 2,791 were excluded, 621 full-text articles were assessed, and 601 were excluded, leaving 20 studies for thematic synthesis. Sixteen were primary empirical or applied studies, while four were secondary evidence sources, including two scoping reviews and two other review articles, which the authors analysed separately to avoid double-counting findings.

The included studies ranged from randomised controlled trials and pre–post evaluations to qualitative interviews, surveys and technical model development papers, spanning primary care, pediatrics, cardiology, mental health and chronic disease management. Five themes emerged: high diagnostic and predictive performance of machine learning models; enhanced clinical decision-making and workflow efficiency; the importance of user trust and acceptance; the role of explainability in improving interpretability; and significant implementation barriers, including infrastructure and regulatory challenges.

On performance, the review found consistently strong diagnostic accuracy and predictive capability across clinical domains. In cardiology, Elvas and colleagues trained predictive models on patient-level clinical data to identify early signs of myocardial ischemia, achieving high sensitivity and specificity. In pediatrics, Ramgopal and colleagues reported strong discrimination, with high AUC values, using large-scale pediatric datasets for risk prediction. In infectious disease, Mayya and colleagues built a deep learning-based system for COVID-19 diagnosis using multimodal patient data, achieving high diagnostic accuracy. The authors caution, however, that many studies report high internal validation metrics while fewer provide external validation across different populations, and that inconsistent reporting of accuracy, sensitivity, specificity and AUC complicates direct comparison between studies.

Clinical utility emerged as more than accuracy. The review distinguished systems that meaningfully alter clinician behaviour from passive information tools. Randomised trials carried the strongest evidence: Moja and colleagues reported that a hospital-based decision support system improved clinicians’ adherence to guideline-recommended practices, while McGinn and colleagues showed that an evidence-based system in primary care enhanced diagnostic decision-making and increased appropriate guideline use. Applied studies suggested machine learning-driven systems could optimise medication selection by analysing patient-specific variables and assist treatment selection in mental health by integrating clinical history, symptom patterns and treatment responses. Yet the review stresses that these benefits depend on seamless workflow integration; poorly aligned systems can disrupt workflows and add cognitive burden, and clinicians risk over-reliance on algorithmic recommendations.

Adoption and trust proved decisive. Survey and qualitative studies indicated that clinicians’ willingness to adopt these systems is shaped by perceived usefulness, ease of use and confidence in algorithmic outputs. Zhao and colleagues found that intention to adopt AI-based systems was strongly influenced by perceived usefulness and ease of use, shaped by prior experience, digital literacy and familiarity with AI. Qualitative work by Giebel and colleagues identified trust as a central determinant of uptake, with clinicians wary of opaque algorithms and erroneous outputs, particularly in high-stakes settings; confidence rose when systems provided clear justifications for recommendations. Wang and colleagues, studying deployment in rural clinics, documented how limited infrastructure, inadequate training and workflow incompatibilities hinder adoption even where benefits are recognised. Clinicians also expressed concern about automation bias and ambiguity over liability when algorithmic recommendations influence decisions.

Explainability was identified as a key requirement for trust and interpretability. Zhang and colleagues developed an explainable AI-driven system showing that mechanisms such as attention maps and feature attribution can improve clinicians’ understanding of how predictions are generated. Chan and colleagues, in a simulation-based evaluation of an AI-based system for gastrointestinal bleeding risk, found that embedded explanatory components helped clinicians contextualise risk predictions, but also noted that visualisations do not always align with how clinicians conceptualise clinical evidence. The review highlights a deeper tension: explanations generated by machine learning models are often post hoc approximations rather than true representations of underlying decision processes, and current explainable AI techniques only partially bridge the gap between technical transparency and clinical reasoning.

Implementation barriers were widely reported and interconnected. Wang and colleagues showed that limited access to reliable hardware, internet connectivity and electronic health record systems restricts the functionality of AI-driven systems in rural settings, raising equity concerns as benefits accrue disproportionately to well-resourced environments. Giebel and colleagues identified integration challenges from fragmented data systems, where inconsistent data formats and legacy technology hinder the real-time data exchange predictive models require. Resistance to change, insufficient training and lack of adaptability to local clinical contexts further undermined uptake. Scipion and colleagues emphasised regulatory and ethical uncertainties, including data governance, patient privacy and legal accountability, noting that approval processes are often misaligned with the iterative nature of machine learning, which requires continuous updating and validation.

The review’s authors conclude that machine learning-driven decision support offers substantial potential benefits but that successful integration requires addressing trust, explainability and health system readiness. They note the review’s limitations: only English-language literature was included, preprints, conference publications and review articles were eligible alongside peer-reviewed studies, and, consistent with scoping review methodology, no formal risk-of-bias appraisal was undertaken, so the findings map the evidence rather than assess intervention effectiveness. Future research, they argue, should prioritise real-world validation, external validation across diverse populations, human-centred design and context-sensitive implementation strategies, particularly for low- and middle-income settings where digital infrastructure is often underdeveloped.

Subject of Research: Machine learning applications in clinical decision support systems

Article Title: Machine learning models in clinical decision support systems: a scoping review

Article References: Innocent, D. C., Innocent, R. C., & Innocent, I. P. (2026). Machine learning models in clinical decision support systems: a scoping review. Bulletin of the National Research Centre, 50(1), Article 115. https://doi.org/10.1186/s42269-026-01502-y

Image Credits: AI Generated

DOI: 10.1186/s42269-026-01502-y

Keywords: machine learning, clinical decision support systems, scoping review, explainable AI, health informatics, diagnostic accuracy, clinical adoption, implementation barriers, health system readiness, artificial intelligence, clinician trust, workflow integration