Machine Learning Triage Tools Reach the ED, but Their Interfaces Still Look Underbuilt
Machine learning has moved into emergency department triage, yet the user interface work around these systems has lagged behind the model development itself. A systematic review of machine learning-based clinical decision support systems for routine emergency department triage found that adoption remains constrained not just by prediction quality, but by how the tools present risk, uncertainty, and next steps to clinicians.
The review screened 2,678 records and included 25 studies published between 2015 and 19 August 2025. Most of the included work was recent: 80% appeared in the last three years. Even with that momentum, only 36% of the systems were integrated into the healthcare workflow at the time of study. In a setting defined by rapid decisions, crowded worklists, and constant cognitive load, interface design can determine whether an algorithm becomes part of care or remains a demonstration project.
One sentence captures the central problem. Most systems were built to predict, not to be used.
What the Review Found in the Design of Triage Interfaces
The investigators evaluated system characteristics, interface decisions, and design processes across the eligible studies. Almost two-thirds of the systems used a visual component, but none used audio alerts. In practice, that means the interface often depended on display-based communication alone, even in a department where attention is fragmented and clinicians may be moving between patients, monitors, and documentation screens.
The most striking pattern emerged in sub-problem triage systems. Eight of 10 presented only a risk score or binary indicator, without accompanying recommendations or guidance for action. A score can identify danger, but it does not explain the bedside response. In emergency care, that gap matters because triage decisions are not abstract classifications; they trigger bed allocation, escalation, observation, and resource deployment.
Explainability features were not rare, yet uncertainty and model performance indicators appeared infrequently. That is a consequential omission. A clinician asked to trust a model needs more than a label, especially when pretest probability shifts minute by minute and when the cost of a false reassurance can be high. Transparent uncertainty reporting can help clinicians calibrate trust, while performance context can reduce overreliance on a system whose behavior may differ across patient populations and workflows.
Why Interface Design Shapes Whether Algorithms Change Care
The review places interface design at the center of implementation failure rather than at the margins. Emergency department triage is a high-stakes cognitive environment, and machine learning output enters that environment as one more input competing for attention. If the output is buried, ambiguous, or stripped of actionability, the tool may be technically sound and operationally weak at the same time.
User-centered design has long been tied to adoption in digital health, and the same logic applies here. A triage model that predicts risk without indicating what to do next asks clinicians to perform an extra layer of translation under pressure. That can slow decisions, create uncertainty about responsibility, and make the tool feel like surveillance rather than support. Interface choices are not cosmetic in this setting; they determine whether a signal becomes a decision aid.
Commercial systems stood out in the review. Their interfaces were more elaborate, but they were less likely to be integrated at the time of publication. That pattern suggests a familiar split between polish and penetration. A refined interface can make a product appear mature, yet without embedding into clinical workflow, the system still stops short of routine use. Integration depends on fit with local processes, staffing, and the rhythms of triage, not just on visual design.
Sparse Testing and Limited Reporting Leave the Field With Weak Comparators
Very few studies described a structured approach to design or testing. Several authors attributed shortcomings of their systems to interface-related problems, but the underlying methods for identifying and correcting those problems were rarely reported in detail. That leaves the field with a familiar and frustrating evidence gap: developers recognize the interface as a weak point, but the published record often fails to show how they tried to fix it.
The descriptive nature of the review also reflects a broader problem in the literature. Heterogeneous study designs and limited reporting prevented a more definitive synthesis. Rather than proving that one interface pattern is superior, the review maps the current state of practice and shows how little of it is grounded in reproducible human-computer interaction methods. For a domain built on clinical prediction, the evidence base for presentation and interaction is still thin.
This is not a minor implementation detail. In emergency medicine, triage tools compete with alarms, handoffs, charting demands, and urgent bedside assessments. An interface that fails to communicate risk clearly, or one that omits uncertainty and next-step guidance, may contribute to alert fatigue or passive disregard. A model that cannot earn a clinician’s attention at the point of care cannot improve throughput, prioritization, or safety in any reliable way.
Interdisciplinary Design May Decide Whether Triage AI Scales
The review points toward a practical path forward: systematic design processes, interdisciplinary collaboration, stronger use of human-computer interaction principles, and transparent reporting. Those elements are not decorative additions to machine learning triage; they are the conditions that make such tools usable in real departments. Clinicians, informaticians, designers, and implementation specialists need to work from the same interface logic, not from separate assumptions about what the model should say and what the user needs to see.
The setting matters as much as the software. Emergency triage is a high cognitive-load environment where small interface failures can have outsized consequences. A color cue, a missing recommendation, or an unexplained score can alter behavior at exactly the wrong moment. The review makes plain that the implementation gap in machine learning triage is not only about algorithmic accuracy. It is also about whether the interface can translate prediction into action with enough clarity, speed, and credibility to matter at the bedside.
Future evidence will need to do more than report model performance. It will need to describe how interfaces were designed, who participated in testing, what uncertainty signals were shown, and how the tool behaved once inserted into the triage workflow. Without that level of reporting, comparisons across systems will remain difficult, and the field will keep rebuilding the same lessons in different software shells.