User interfaces in machine learning-based decision support for emergency department triage — a systematic review

Background Triage is a promising application for machine learning-based clinical decision support systems (CDSSs). However, despite considerable research attention, their real-world impact remains limited, with persistent challenges in implementation and adoption. While user-centred interface design and testing are known to improve adoption and outcomes, if and how they are considered in real-world systems remains understudied. Methods We conducted a systematic review (PROSPERO ID: CRD420251119077, funded by affiliated institutions 1,2 ) to characterize and descriptively analyse interface design characteristics and processes for machine learning-based CDSSs for routine emergency department triage. To this end, PubMed, Embase, CENTRAL, Web of Science, and IEEE Xplore were searched for papers published between 2015 and 19 August 2025. Two reviewers independently screened all records and extracted and coded data on system characteristics, interface design decisions, and design processes. Results Of the 2,678 records screened, 25 studies met the inclusion criteria. Of these, 80% had been published in the last three years, and 36% of systems were integrated into the respective healthcare workflow at the time of the study. Almost two-thirds of systems used a visual component, but none used audio alerts. In the case of sub-problem triage, 8 of 10 systems presented only a risk score or binary indicator, without accompanying recommendations or guidance for action. Although explainability features were not uncommon, uncertainty and model performance indicators were rarely used. Overall, commercial systems had more elaborate interfaces but were less likely to be integrated at the time of publication. Very few studies describe a structured approach to their design or testing, despite several authors attributing their systems’ shortcomings to interface-related issues. Conclusion Despite the established importance of user-centric design and testing in CDSS development, these practices remain limited in ML-based clinical decision support tools for triage. Due to the heterogeneity of studies, descriptive approach and limited reporting, this review only presents a first step towards the investigation of this important lever. Especially in high cognitive-load settings such as the emergency department, closing the ML implementation gap will require systematic design approaches, interdisciplinary collaboration, stronger integration of human–computer interaction principles, and transparent reporting to improve accountability, comparability, and future evidence synthesis. Systematic Review Registration https://www.crd.york.ac.uk/PROSPERO/view/CRD420251119077 , PROSPERO CRD420251119077.