Featured case study · AI-assisted healthcare

Designing human oversight for real-time AI alerts.

Sentinel E-Sitter helped remote patient observers monitor several rooms, assess fall-risk alerts, intervene with patients, and coordinate an in-room response.

RoleLead UX Designer
PlatformDesktop browser
TeamFrontend, backend, QA
OutcomeNine-room pilot

The challenge

Urgency without chaos.

One trained e-sitter could monitor up to nine patient rooms during the pilot. Several patients could trigger alerts at the same time, so the interface needed to make risk visible without turning every feed into competing noise.

The AI could detect sitting up as medium risk and leaving the bed as high risk. It did not replace judgment: the e-sitter still needed to validate the alert, begin an intervention, contact a nurse, and document the event through resolution.

Core workflow

From detection to a human response.

1 · Detect

AI identifies medium- or high-risk behavior across live feeds.

2 · Validate

The e-sitter confirms or rejects the alert and starts intervention.

3 · Coordinate

The observer documents the call and a nurse-response timer begins.

4 · Resolve

A visible status lets the observer confirm arrival and close the event.

Attention hierarchy

Using color and sound with restraint.

Monitoring

No-risk feeds remained unhighlighted to reduce distraction.

Medium risk

Yellow identified a patient sitting up and requiring attention.

High risk

Red and an audible alert signaled a patient leaving the bed.

Six Sentinel monitoring views showing audio intervention, AI approval, high-risk rejection, patient details, simultaneous yellow and red alerts, and a focused room view.
The shared monitoring view supported intervention, alert validation, parallel risk states, and a focused patient view without losing awareness of the other rooms.
Sentinel interface details showing patient history, high-risk alert validation, live audio intervention, privacy screen controls, and a speaker error message.
Workflow details included patient context and history, explicit alert validation, live audio state, privacy controls, and visible equipment feedback.

Human control

The system suggested; the observer decided.

Thumbs-up and thumbs-down controls allowed e-sitters to validate AI output. Rejecting an incorrect alert produced an immediate confirmation, and the feedback contributed to improving future detection.

For high-risk events, the system could suggest a nurse from its database while preserving the flexibility to add someone manually. A status pill over the patient feed let the e-sitter confirm that the nurse had arrived.

Testing and iteration

Making the next action easier to understand.

Audio intervention

Unclear icon affordances gained explanatory labels on hover.

Required information

Inputs were revised to make missing information more accessible and visible.

Expiring contacts

A coordinated badge and sidebar countdown made the time-sensitive state clear.

AI feedback

An onboarding guide explained how validation controls helped train the system.

Adding a contact

The action became more prominent and easier to find in the required flow.

Evidence

Approximately eight participants evaluated Figma prototypes and the workflow.

Reflection

What I would take further.

Automatic movement of urgent rooms into a prioritized position was identified as a valuable next step, but it was not implemented before budget constraints ended the contractor engagement. I would also explore giving trained users clearer context for why an alert fired—without adding noise or implying certainty the model could not support.