Healthcare UX: Where Clarity Becomes Clinical Safety

In most software, a confusing screen costs a few seconds or a lost conversion. In clinical software it can cost something else. A medication list that truncates the dose, a default that pre-selects the wrong unit, a patient banner that scrolls out of view while an order is entered: these are ordinary interface decisions, and in a clinical setting each one can contribute to an error.

That is not a hypothetical concern. Howe, Adams, Hettinger, and Ratwani (JAMA, 2018) reviewed about 1.7 million patient-safety event reports from 571 Pennsylvania healthcare facilities and a large mid-Atlantic health system, covering 2013 to 2016. They identified 557 reports in which language suggested EHR usability contributed to possible patient harm. Most required monitoring to prevent harm, 14% potentially caused temporary harm, and a small number were more serious. Because the study only counted reports that named a specific vendor or product, and safety events are underreported, the authors note the real figure is likely higher.

In healthcare, clarity is not a visual preference. It is part of how the system prevents error.

Usability problems also show up as time. In a time-and-motion study of 57 physicians across family medicine, internal medicine, cardiology, and orthopedics, Sinsky et al. (Annals of Internal Medicine, 2016) found that physicians spent 27% of their day on direct face time with patients and 49% on EHR and desk work. That is roughly 2 hours of screen and desk work for every hour with patients. Diaries kept by a subset of physicians showed another 1 to 2 hours of computer work most nights.

Design does not explain all of that time. Documentation requirements, billing, and policy drive much of it. But interface decisions determine how much effort each requirement takes. Workflows that require re-entering information, navigating between disconnected screens, or hunting for a result are design problems with a direct cost in clinician time and attention.

Clinical decision support is supposed to catch errors. In practice, much of it is ignored. A review of 17 studies by van der Sijs and colleagues (JAMIA, 2006) found clinicians overrode drug-safety alerts in 49% to 96% of cases. Many overrides were appropriate, because the alerts were not specific to the patient or the situation. The result is a system that trains people to dismiss warnings, including the rare one that matters.

This is a design problem as much as a clinical-content problem. Alerts that interrupt for low-risk issues, use the same visual weight for every severity, and give no context for why they fired all push clinicians toward reflexive dismissal. Fewer, more specific, and better-placed alerts protect the ones that matter.

AI-assisted decision support adds a new layer. The FDA's final Clinical Decision Support Software guidance (September 2022) explicitly discusses automation bias, describing it as the tendency to over-rely on a suggestion from an automated system. One of its criteria for software that is not regulated as a device is that a clinician can independently review the basis for the recommendation, rather than relying primarily on it. The guidance recommends disclosing the intended use, intended users and patients, required inputs, and a plain-language description of how the algorithm was developed and validated. Software intended for time-critical decisions does not meet that criterion.

Whatever the regulatory path, the design implication is the same. Clinical AI interfaces should show what the recommendation is based on, what it does not know, and how the clinician can confirm or override it, at the point where the decision happens. Research on automation more broadly supports this: Parasuraman and Manzey (2010) found that experts are as susceptible to automation bias as novices, and that training alone does not prevent it. The interface has to do some of the work.

Make the patient context impossible to lose

Patient identity, allergies, and critical context should stay visible during every consequential action. Wrong-patient errors are among the most preventable, and persistent identification is one of the simplest defenses.

Design for interruption

Clinicians are interrupted constantly. Workflows should save state automatically, show clearly where the person left off, and never leave a half-finished order in an ambiguous state.

Separate signal from noise

Use visual weight and interruption only for what truly needs it. Tier alerts by severity, make low-risk information passive, and give every interruptive alert a reason the clinician can see in one glance.

Prevent errors instead of reporting them

Constrain inputs where it is safe to do so: sensible units, dose ranges, and defaults that match the most common safe case. Prevention in the interface is cheaper and safer than an alert after the fact.

Test with real roles in realistic conditions

The FDA's human factors guidance for medical devices centers on identifying use-related hazards and validating critical tasks with representative users. Even products outside device regulation benefit from that discipline: test critical tasks with physicians, nurses, lab staff, and patients separately, because each role reads the same screen differently.

1. During any order or change, is the patient's identity and critical context always visible?
2. What share of your alerts are overridden, and do you know why?
3. Can a clinician resume an interrupted task without re-reading or re-entering anything?
4. For AI recommendations, can the clinician see the basis, the inputs used, and the limits, in the same view?
5. Have the top five critical tasks been tested with each role that performs them?
6. How many clicks and screens does the most common daily task take today?

Healthcare products are judged by a harder standard than most software, and they should be. The evidence is consistent: interface decisions affect safety, clinician time, and whether safety systems are trusted at all. Treating clarity as part of clinical safety, rather than as surface polish, is the most practical way to build products clinicians can rely on, and that patients are safer for.