Designing AI for Breast Cancer Risk Communication
DURATION
May - Nov 2025
ROLE
Conversational Designer
Product Researcher
Tools
Dialog Flow
Figma
COLLABORATION
Machine Learning Engineer
TL;DR
Breast cancer risk models like Gail and Tyrer–Cuzick are widely used to estimate women’s 5-year and lifetime cancer risk. They are statistically rigorous—but emotionally complex. When risk scores are shared without context, patients may experience panic, false reassurance, or avoidance—especially in a disease space tied to identity, family history, and long-term life planning.
This project asks a practical question: where should AI sit in the real breast cancer care journey—before the visit, during triage, or after results? Rather than redesigning the model itself, we redesigned how risk is introduced, paced, and framed so it can support understanding without escalating anxiety or overstepping clinical boundaries.
I designed a Dialogflow-based conversational prototype plus a clinician-aligned summary output. The system integrates emotional readiness checks, insurance clarification, gradual score disclosure, and structured reporting—positioning AI as a preparation layer for medical conversations, not a replacement for them.
Challenge
Risk scores are precise—but understanding them is not.
Breast cancer risk models such as Gail and Tyrer–Cuzick are two of the most widely used tools in clinical practice for estimating 5-year and lifetime cancer risk. While statistically robust, these models were originally designed to support clinician decision-making—not to independently guide patients.
In practice, however, risk scores are often communicated directly to patients without structured support. This creates a structural tension:
A low-risk result may create misplaced reassurance
A high-risk result may feel like a verdict
Cost and insurance concerns may quietly discourage follow-up
AI explanations may be over-trusted—or dismissed—depending on context
The challenge was not to redesign the statistical model, but to redesign how risk is disclosed in an AI-mediated setting.
Research
Understanding risk communication across the care journey
Patient & Clinician Interviews
I spoke with 5+ potential patients (low/high risk) and Dr. Nina to understand how risk scores are received in real life. I found that the biggest failure points are timing, cost/insurance uncertainty, and the lack of a clear next-step framing—more than the score itself.
Stakeholder Action Log
Across the care journey, risk interpretation often falls on patients in moments of limited time and fragmented communication. Coverage, urgency, and explanation—not just medical data—shape whether women act on their risk.
User Journey Map
Rather than inserting AI randomly into the flow, this map helped determine where conversational support could reduce cognitive and emotional burden without disrupting clinical roles
Strategy
Based on my research, this strategy followed four principles:
01 Ask readiness before showing risk
02 Clarify insurance & cost early
03 Show risk in steps (not one raw number)
04 Generate a visit-ready summary for the clinician
Solution
I designed a Dialogflow-based conversational system that translates these principles into structured dialogue logic.
Conclusion
This project reshaped how I think about designing AI in medical contexts. I learned that the difficulty of risk communication is rarely about the accuracy of the model—it is about timing, framing, and the emotional state of the person receiving the information.
Through mapping stakeholder roles and real care workflows, I began to see AI not as a decision-maker, but as a structural layer within a larger system. Designing the conversation required understanding clinical constraints, insurance realities, and cultural factors—not just prompt logic.
Most importantly, this work reinforced that in sensitive health domains, the goal is not to make AI more persuasive, but to make it more responsible.