EVA · Case study · 2026

An AI-assisted risk dashboard for early ovarian cancer detection.

4-month capstone. I designed and researched an AI risk-scoring dashboard clinicians could actually trust.

Design & Research4 monthsTeam of 4Clinician web app

EVA results screen showing a 75% risk score and symptom cards grouped by body system

01 — Problem

70% of ovarian cancers are diagnosed at Stage III–IV.

Symptoms are vague alone. Nothing forces them into one pattern.

70%

diagnosed at Stage III–IV

5 of 8

daily clinician hours in the EHR

44%

of that time scanning for a pattern

02 — Research

What earns trust in an AI-generated score.

Two numbers from secondary research set hard constraints on the design.

87% → 33%

override rate, without vs. with visible reasoning

Decision — every score ships with its reasoning attached.

44%

of EHR time spent scanning notes for a pattern

Decision — replaced tables and text with graphical summaries.

Existing physician practice dashboard showing Meaningful Use quality measures, examined during competitive research — Existing tools in the market — built for billing compliance, not for spotting a pattern.

03 — What clinicians pushed back on

Interviews and usability testing surfaced three tensions.

Each shaped a specific decision in the final design.

Primary interview

Clinician A

"If we don't encourage self-googled or ChatGPT diagnosis — how can our patients expect the same?"

Finding: wants AI to inform judgment, not replace it.

Usability testing

Testing cohort

"Minimal, but want all the important information at the same time."

Finding: brevity and completeness, both required at once.

V1 feedback

Clinician reviewer

"Too much purple."

Finding: visual tone read as unserious for clinical use.

04 — Before / after

Same content, restructured from tables into charts.

Information density didn't change. How it's read did.

Before

Earlier version of the EVA results screen with a purple theme, ESGO 2024 protocol reference and a numbered list of triggered rules

Results read as a compliance checklist.

After

Final EVA results screen with a 75% risk score ring and symptom cards grouped by body system

Results read as a clinical picture.

Before

Earlier version of EVA explainability presenting findings as a list of collapsible cards with a purple theme

Findings listed one below another.

After

Final EVA explainability screen showing findings as a radar chart with a short structured red-flag statement

Same findings, as a chart and one red flag.

05 — Supporting process

Structure and navigation, tested before visual design.

Mid-fidelity wireframe used in moderated usability testing, showing key findings as structured placeholder cards — Mid-fi wireframe, moderated usability testing.

Mid-fidelity wireframe used in moderated usability testing, showing a risk score summary layout — Mid-fi wireframe — risk score layout variant.

Doctor's non-linear clinical decision journey, mapped as a flowchart with pre-consultation and doctor's-perspective swim lanes — Journey mapping — the path loops, not linear.

Hand-drawn sketches exploring risk score gauge and dial variations — Chart-type sketches, tested for credibility.

06 — Final design

Four screens, each tied to a decision above.

Applies: evidence attached to every score

01 — Results

Symptoms grouped by body system; confounders surfaced up front.

Final EVA results screen with risk score and symptom cards grouped by body system

Applies: reasoning visible, not just a score

02 — Explainability

A radar chart makes cross-system overlap hard to dismiss.

Final EVA explainability screen with a radar chart showing symptom overlap

Applies: nothing recommended without a reason

03 — Recommendations

Urgency tiering, with rationale and timing for every action.

Final EVA recommendations screen with a timeline of urgent and near-term actions

Applies: one clean, clinician-grade handoff

04 — Export

A structured report ready for a specialist or the patient record.

Final EVA export screen with a downloadable clinician-grade report preview

07 — Reflection

What shipped, and what I'd test next.

Every element needed a research-backed reason.

Collaboration meant defending research, not just agreeing.

Deciding what not to show was the real work.

Next: A/B test the layout, track decision time post-launch, stress-test edge cases.