Rescuing the Read: Overhauling MederiAI's Endoscopy Reviewer
Medical AI · Capsule endoscopy | MederiAI | Reviewer redesign
Summary
A reviewer where the AI and the doctor fought for the screen
MederiAI builds AI for capsule endoscopy — a pill-sized camera that travels the GI tract and captures tens of thousands of frames the model scans for findings. The model was strong. The reviewer screen doctors used to read its results was not: a dense, three-column wall that buried the clinician in everything the AI saw. I was brought in mid-project to rescue the read.
This case study is one reframing: the AI shouldn't lead the diagnosis — it should triage for a doctor who does. Everything else followed from that.
Product
Web · Clinician-facing capsule-endoscopy reviewer
Timeline
2024 · ~2 weeks (2 iterations)
My Role
Sole UI/UX designer. I ran a heuristic evaluation of the live reviewer, turned the findings into prioritized fixes, and rebuilt the read around the clinician's judgment — the AI as triage, never the authority.
Skills
Information Architecture, Interaction Design, Data-Dense UI, Heuristic Audit, Clinical UX, AI-assist Patterns
Impact
Shipped to production — ~83% fewer clicks to action an AI finding (6→1), plus the undo, review-progress and dark mode the old reviewer lacked. Validated by practicing gastroenterologists and the founding team.
Roles & Research
A few roles touch this reviewer. I built on and enhanced the team's existing user-journey map, then focused the redesign on the one carrying the risk — the reviewing clinician.
I built on the team's existing user-journey map — extending and enhancing it rather than starting from scratch. A few readers, one protagonist: the reviewing clinician, confirming or dismissing AI findings under time pressure, is who this redesign solves for. Here are the readers I designed for:
I started from the team's existing user-journey map and extended it — refining the reviewing clinician's path where the read needed it — rather than running new field research. My core method was a heuristic evaluation: I walked the live reviewer against usability heuristics and the logic of a clinical read, ranking every place it added cognitive load or took the call away from the doctor. The clinicians came in afterwards — to test the redesign.
The Problem
"It shows me everything — which means it shows me nothing."
The reviewer surfaced every AI detection at once, across three dense columns, with the model's calls visually louder than the doctor's. Triaging a study meant fighting the layout: hunting for the findings that mattered, second-guessing the ones that didn't, and never quite feeling in control of the diagnosis.
An assistant that overwhelms isn't assisting — it's competing.
The model was confident — the screen made the doctor doubt it.
How I Worked
Audit first, then rebuild the read
I don't polish a clinical screen from the outside — I evaluate it against how the read actually has to work, then design against the friction. On MederiAI it looked like this:
The Wrong Assumption
"Show the doctor everything the AI sees."
The original reviewer treated completeness as safety: surface every detection so nothing is missed. But for a doctor mid-read, "everything, equally loud" is the same as noise — the critical finding sits beside fifty trivial ones, all demanding the same attention. Completeness without triage doesn't reduce risk; it relocates it onto a tired clinician.
Safety isn't showing everything — it's surfacing the right thing first.
A critical finding, loud beside fifty trivial ones, is a finding you'll miss.
How I Decided
Optimising for the doctor's judgment, not the model's output
I weighed three ways to present a strong model to a clinician who has to own the call.
Approach | Doctor in control | Surfaces critical first | Risk |
|---|---|---|---|
Tidy 3 columns | No | No | High |
AI auto-verdict | No | Yes | High |
Two-pane, triaged | Yes | Yes | Low |
I optimised for the doctor's judgment, not the model's confidence.
The Redesign
Wireframes & Structure
Before any dark-theme polish, I re-architected the reviewer in low fidelity — proving the structure worked before styling it. The core move: collapse three competing zones into one focused, reading-first layout.
The low-fi proved four moves at once: the data-entry form became an on-demand modal, findings became a severity-ordered rail, lesion position moved onto a real timeline, and four competing CTAs collapsed into one path.
A focused, dark, two-pane read
I collapsed three columns into a calm two-pane read built for long, careful sessions.
Severity-triaged findings
AI detections are ranked Normal / Warning / Critical, with severity carried by icon + label, never colour alone — so the worst finding is the first one the doctor meets.
A focused read pane
One frame at a time, large and central, with the AI's mark as a quiet overlay the doctor can trust or dismiss — not a verdict shouting over the image.
A report that writes itself
The report builds as the doctor reads — by region and finding — so writing it is a by-product of the read, not a second job after it.
The doctor always in control
Every AI call is accept / override, logged and reversible.
The model proposes; the clinician disposes.
The strongest model still has to wait for the doctor to look.
Built for the Worst Moment
Designed for the end-of-day read
A reviewer is judged on the twentieth study of a long day, not the first. So I designed for fatigue and the cost of a miss.
Critical can't hide — a Critical finding is surfaced first and can't be scrolled past unacknowledged.
Severity never relies on colour — icon + label, so a tired or colour-blind reader reads it instantly.
Overrides are reversible — a mistaken dismissal is one step back, never a silent loss.
Degraded confidence is honest — low-confidence AI calls are marked as such, never dressed up as certainty.
I treated the tired clinician on the hardest case as the real user — because that's the moment the screen has to hold.
Impact & Learnings
Validated by the people who'd use it
Tested with practicing gastroenterologists and MederiAI's founding team — a CEO with a PhD in Robotics, AI & Control Theory (DTU) and a CTO with an MSc from Aalborg — and shipped to production.
“This will make our workflow seamless.”
— MederiAI founding team
Outcomes (qualitative)
A read you can sustain — the two-pane layout holds up across a long session, not just a demo.
AI as assistant, not authority — the model's strength shows without taking the doctor's call.
Trust restored — clinicians stopped fighting the screen and started using the AI.
Learnings
Completeness is not safety. Triage is.
In a diagnostic tool, the UI's job is to protect the clinician's judgment, not replace it.
Design for the tired user on the hardest case — the easy ones take care of themselves.
An AI a doctor trusts is one that knows its place — assisting, never deciding.
Clinicians stopped fighting the screen and started trusting the AI.
What's Next
Close the loop with the ML team — feed override patterns back to sharpen model confidence.
Extend the read pattern — beyond capsule endoscopy to other study types.
Tune the report with referring physicians — so the output reads as cleanly as the input.
Designing AI that respects the expert — one read at a time.
Good medical AI knows exactly where its job ends.