PortfolioCase StudiesClinical Report Intelligence
✶ HealthTech · AI Clinical Decision Support · Trust Design · Conceptual Project

Clinical Report
Intelligence Platform

Designing an AI decision-support layer that clinicians actually trust — by making reasoning visible, not hidden.

In clinical genomics, the challenge is not access to data — it is decision-making. Reports are dense, decisions are high-stakes. Raw data increases cognitive load. AI outputs without transparency reduce trust. This creates a gap between information and action.

Role
UX Designer
Domain
Healthcare AI · Clinical Genomics
Focus
AI Trust · Decision Support
Type
Conceptual / Speculative Design
Users
Genome Analysts · Clinicians
Clinical Report Intelligence Platform cover
01 — Context

The challenge is not data access.
It is decision-making.

In clinical genomics, reports are dense, decisions are high-stakes, and validation requires cross-referencing multiple sources. Most existing tools either expose raw data or attempt to simplify through AI. Both approaches introduce friction.

01
Raw data increases cognitive load

Clinicians working with unstructured genomic reports must manually extract signals, prioritise risks, and cross-reference databases — all under time pressure and with patient outcomes at stake.

02
AI outputs without transparency reduce trust

Systems that deliver AI conclusions without showing reasoning create a different problem: clinicians cannot evaluate the output, so they either over-rely on it or reject it entirely.

03
The gap between information and action

Neither raw data nor opaque AI closes this gap. What clinicians need is a system that structures information around how decisions actually get made — iteratively, with evidence, not answers.

02 — Understanding the User

Genome analysts and clinicians
working with complex diagnostic data.

This workflow is used by genome analysts and clinicians who process high-volume, high-stakes genomic data. A few consistent patterns shaped every design decision.

Decisions are iterative, not linear

Clinicians orient, prioritise, inspect, compare, then act — and often cycle back. Rigid sequences break natural reasoning at the worst moment.

Users look for evidence, not answers

Clinicians do not want an AI to deliver a conclusion. They want evidence they can evaluate and stand behind. Confidence comes from corroboration, not a single metric.

Confidence comes from corroboration

A single percentage score is not enough. Clinicians need to see the basis for an AI assessment — and assess whether the evidence is strong enough to act on.

Systems that remove control are rarely adopted

Regardless of efficiency gains, tools that reduce clinician agency create distrust. The design position: AI as an interpretable assistive layer that supports judgment — never replaces it.

The framing shift: This was approached as a decision-support system, not a visualisation problem. The key question: “How might clinicians arrive at confident, defensible decisions with AI support?”
03 — Key Observation

The critical moment is not when AI produces a result.
It is when the user asks: “Do I trust this?”

Most systems do not design for this moment. This became the central design problem.

Position: AI in healthcare should not act as an authority. It should function as an interpretable, assistive layer that supports human judgment.

Every design decision was evaluated against one principle: Does this help the user make a better decision?

The focus remained on: reducing ambiguity, supporting validation, and making system behaviour understandable.

04 — System Flow

Structured around how
decisions naturally evolve.

The system is structured around how decisions naturally evolve — avoiding rigid workflows and supporting natural reasoning at every stage.

1. Upload report and patient data

Minimal friction intake. The system orients the clinician to the case before beginning interpretation.

2. AI analyses and identifies signals

The system processes the report and surfaces key genomic signals, prioritised by risk level.

3. Summary highlights key risks

A structured summary tells the clinician what needs attention first — without requiring them to read the full report.

4. User inspects evidence

Clinicians drill into individual variants, reviewing supporting evidence from ClinVar, COSMIC, gnomAD, and literature.

5. AI reasoning is explored

Full AI reasoning is available on demand — signal strength, evidence sources, consistency, and validation gaps.

6. AI and human interpretations are compared

A dedicated comparison view surfaces where AI and clinician assessments agree, partially agree, or diverge.

7. User validates or overrides

Overrides are a first-class action. Clinician reasoning is explicitly captured. AI output remains visible throughout.

8. Final report is generated

Every decision step is traceable in the output. The report is defensible, not just generated.

05 — Dashboard

Orientation and prioritisation
before the clinician reads a word.

The dashboard gives clinicians an immediate case-level view — total reports, pending reviews, AI accuracy, and active patients — with recent cases listed by risk level. The clinician knows what needs attention before opening a single report.

Clinical genomics dashboard — case list with risk levels and AI accuracy metrics AI processing screen — analysis in progress Upload report and patient data screen

The information architecture reflects real decision patterns: Orientation → What is this case? · Prioritisation → What needs attention? · Inspection → Why is this flagged? · Comparison → Do I agree? · Action → What decision do I make?

06 — Key Design Decisions

Five decisions that define
how trust is built.

Decision 01
Making AI Reasoning Visible

Hiding reasoning reduces trust. Showing everything simultaneously increases cognitive load. The system uses progressive disclosure: summary by default, reasoning on demand. Clinicians access the depth they need, when they need it.

Each variant card shows an AI-generated clinical summary with confidence indicators. Full reasoning — evidence sources, signal strength, consistency, and validation gaps — expands on demand.
AI Analysis Summary screen showing identified variants with confidence levels
Decision 02
Reframing Confidence — Evidence Over Percentage

A single confidence percentage is not actionable. “87% confident” tells a clinician nothing they can evaluate. Confidence is presented as three readable dimensions: strength of evidence, consistency of signals, and need for additional validation.

Clinicians shift from passively accepting a score to actively evaluating its basis — which is how clinical judgment actually works.
Variant detail screen showing BRCA1 evidence breakdown with confidence dimensions
Decision 03
Supporting Critical Evaluation — AI vs. Human Comparison

AI is positioned as input, not decision. A dedicated comparison layer shows AI interpretation and human interpretation side by side, with differences clearly surfaced. Clinicians evaluate where they agree with the system and where they diverge.

Transparency about disagreement builds more confidence than false agreement would. This is the most trust-building feature in the system.
AI vs Human comparison view showing full agreement, partial agreement, and disagreement across variants
Decision 04
Preserving Decision Integrity — Overrides as Expected, Not Exceptional

Most AI systems treat overrides as edge cases — buried, requiring extra confirmation, making the user feel like they are correcting the system. This platform treats overrides as a normal, first-class clinical action. AI output remains visible. User decisions are explicit. Reasoning is captured.

The clinician is never made to feel like they are overruling the AI. They are doing their job, and the system supports that.
Human Override screen showing clinician reclassification of a BRCA2 variant with documented reasoning
Decision 05
Closing the Loop — Corrections as Structured Feedback

User corrections are captured as structured feedback that feeds back into the system’s learning layer. Clinical expertise does not get siloed from the tool that should benefit from it.

Clinicians feel their judgment has weight. The system improves. Trust accumulates over time — not just within a single session.
Final Clinical Report showing validated variant results with AI and human classification side by side
07 — What Was Avoided

Design decisions are also
what you choose not to do.

Over-automation without visibility

No step is automated without the clinician understanding what the system did and why. Faster to demo, dangerous in practice.

Hiding uncertainty

Low signal strength, conflicting evidence, and validation gaps are displayed — not masked. A confident-looking output built on weak evidence is more dangerous than an honest uncertain one.

Reducing complex decisions to single scores

A single percentage transfers accountability to a number that cannot hold it. Confidence is multidimensional and presented that way throughout the system.

Forcing rigid, linear workflows

Clinical reasoning is non-linear. The system supports jumping between orientation, inspection, and comparison without requiring a fixed sequence.

08 — Outcome

Structure introduced into
a fragmented process.

01
Faster identification of relevant signals

AI-prioritised risk summaries reduced time spent manually scanning reports before knowing where to focus.

02
Clear path from insight to decision

The structured flow — orientation to action — gave clinicians a reproducible process that matched how they already reasoned.

03
Increased confidence in AI-assisted workflows

Clinicians who could see AI reasoning and override freely reported higher confidence in the system — not lower. The ability to disagree is the foundation of trust.

09 — Key Learnings

What this project
reinforced.

01
AI systems fail when they hide uncertainty

Transparency about what the system does not know builds more trust than projecting false confidence. Hidden uncertainty is a future failure point.

02
Trust emerges from transparency and control

The ability to see reasoning and override freely built trust. Systems that take control away, regardless of accuracy, do not get adopted.

03
Explainability is a design responsibility

What to explain, how to explain it, and when to surface it is a UX problem. Designers must own this layer — it cannot be delegated to engineering.

04
Designing AI requires thinking in systems, not screens

Individual screen decisions only make sense in the context of the full decision flow. Designing a single screen well means understanding what comes before and after it.

10 — Reflection

From designing interfaces
to designing systems that support reasoning.

This project reinforced a shift in approach: from designing interfaces, to designing systems that support reasoning, judgment, and accountability.

AI should not replace expertise.
It should make expertise more effective.

More Work

See what else I’ve built.