Megan Gallagher Portfolio

Overview

Solution (Video)

Strategy & Direction

System Architecture

Design Decisions

UI Design

Outcomes

OneSkin • Mobile App Design 2024

Designing an AI decision-support system for early skin cancer triage

ROLE

Lead Product Designer & Strategist on a team of 3 engineers

TIMELINE

5 Days (Hackathon)

SCOPE

As the sole designer, I owned product framing, clinical alignment, and all UX decisions from concept through MVP

OVERVIEW

An AI-assisted triage tool designed to support referral decisions under clinical uncertainty

OneSkin explores a critical gap in AI healthcare products: translating probabilistic model outputs into safe, actionable guidance for real-world clinical decisions.

Rather than diagnosing melanoma, the system provides conservative risk signals and structured next steps designed to support referral decisions when certainty is low: the exact scenario that drives over-referral, system strain, and provider liability.

PROBLEM

Primary care lacks safe decision support when skin cancer risk is uncertain

Skin cancer triage doesn’t fail because clinicians lack detection knowledge. It fails when uncertainty is high and risk is unclear.

Under liability pressure, primary care physicians default to referral when unsure, even when malignancy (cancer) likelihood is low. This creates: system overload, delayed specialist access, and unnecessary procedures.

The missing piece is not better detection, it’s decision support for ambiguous cases.

SOLUTION

A patient-facing experience designed to support better upstream triage

A patient-facing experience that standardizes lesion image capture, communicates conservative risk signals, and supports virtual triage, without presenting diagnoses or false reassurance

STRATEGY & DIRECTION

Designing for real-world clinical decision-making

Conversations with two PCPs and two dermatologists showed that the core issue in skin cancer triage isn’t detection, it’s what to do when certainty is low. This shifted the product from an AI detection tool to a decision-support system.

These insights translated into three product principles that shaped every design decision:

AI supports decisions, it doesn’t make them

No diagnostic labels. The system provides conservative risk tiers and clear next steps to support clinician judgment

AI supports decisions, it doesn’t make them

No diagnostic labels. The system provides conservative risk tiers and clear next steps to support clinician judgment

Patient participation improves access, not authority

Patients contribute structured data to enable earlier triage and remote review, while medical decisions remain provider-led.

Patient participation improves access, not authority

Patients contribute structured data to enable earlier triage and remote review, while medical decisions remain provider-led.

Uncertainty must be explicit and actionable

Low-reliability predictions are withheld. Retake flows prevent ambiguous outputs from becoming misleading guidance.

Uncertainty must be explicit and actionable

Low-reliability predictions are withheld. Retake flows prevent ambiguous outputs from becoming misleading guidance.

SYSTEM ARCHITECTURE

How raw AI predictions turn into safe guidance for real people

The main challenge wasn’t just building an image model, it was making sure the model’s output turned into something people could safely act on.

The system works in three steps:

STEP ONE

The model makes a prediction

After a user takes a photo, the AI looks at the image and estimates how closely it matches known patterns of benign (non-cancerous) or potentially malignant lesions

What is the likelihood of this lesion in this image being benign vs malignant

STEP ONE

The model makes a prediction

After a user takes a photo, the AI looks at the image and estimates how closely it matches known patterns of benign (non-cancerous) or potentially malignant lesions

What is the likelihood of this lesion in this image being benign vs malignant

STEP TWO

The system checks if the prediction is reliable enough

The system looks at factors like image quality and whether the model is operating within known limits.

This prevents blurry, poorly lit, or unclear images from turning into misleading results

STEP TWO

The system checks if the prediction is reliable enough

The system looks at factors like image quality and whether the model is operating within known limits.

This prevents blurry, poorly lit, or unclear images from turning into misleading results

STEP THREE

Translates prediction into safe guidance

Instead of showing probabilities or diagnoses, product rules convert the model’s output into:

A conservative risk category (Low / Medium / High)

A clear next step (monitor, schedule PCP visit)

No diagnostic labels

The goal is not to tell users what the lesion is, but what they should do next

STEP THREE

Translates prediction into safe guidance

Product-level rules translate model outputs into Low / Medium / High risk categories and recommended next steps

The goal is not to tell users what the lesion is, but what they should do next

DESIGN DECISIONS

#1: Designing for interpretation, not reassurance

The results screen carries the highest risk in the entire product. Because of this, the goal of the results screen was not to deliver an answer, but to shape interpretation.

#2: Blocking low-quality signals before they become misleading outputs

The system evaluates confidence before surfacing results

Below a minimum confidence threshold (~70%), the experience pauses
Users are prompted to retake the image rather than receiving a low-quality assessment

Rather than surfacing low-confidence outputs (which users often interpret as “probably fine”), I blocked results entirely to prevent false reassurance

#3: Key Tradeoff: Patient-Facing Entry Point vs Full Provider Workflow

Rather than designing a full provider platform, I deliberately focused on the patient-facing intake and decision-translation layer.

This allowed us to validate the safety of risk framing, confidence gating, and conservative guidance (highest-risk part of the system) before expanding into complex clinical workflows.

Current provider benefit: virtual triage

The current system enables patients to share structured lesion images and risk context directly with providers, supporting remote and asynchronous evaluation.

Patients in rural or underserved areas

Patients in rural or underserved areas

Delayed in-person visits or referrals

Delayed in-person visits or referrals

PCPs making triage decisions without immediate specialist access

PCPs making triage decisions without immediate specialist access

Future Direction: Dedicated Provider Experience

A future provider-facing platform would focus on longitudinal decision support rather than single-point predictions.

Confidence trends over time

Confidence trends over time

Clear escalation signals

Clear escalation signals

EHR-ready summaries for documentation and referrals

EHR-ready summaries for documentation and referrals

Longitudinal lesion history across visits

Longitudinal lesion history across visits

UI DESIGN

Designing the interface to reduce misinterpretation and bias

For this project, I intentionally built on Material Design as the foundation given the 5-day Flutter mobile development.

UI design extended beyond visuals into how information is framed, limited, and communicated because misinterpretation carried the highest risk. The interface was intentionally restrained to reduce cognitive bias and false reassurance.

OUTCOMES

Validated a safe, end-to-end decision-support MVP

Within five days, the team delivered a fully functional mobile MVP supporting image capture, confidence evaluation, conservative risk signaling, and provider sharing

Key Outcomes:

End-to-end validation of a patient-facing triage flow under uncertainty
Clinical review and feedback from two primary care physicians and two dermatologists
Top 5 finalist at Miami Hack Week

Most importantly, the project validated that conservative risk framing, retake flows, and non-diagnostic language were clinically acceptable and aligned with real-world triage decision-making.