The Stakes
When a patient's condition begins to deteriorate in a busy hospital ward, the window for early intervention is narrow. Miss it, and the cost is measured in outcomes: longer recovery, worse prognosis, sometimes a life. Sepsis alone is one of the leading causes of ICU mortality, and it's notoriously difficult to catch early enough to change outcomes.
Peach IntelliHealth, developed in collaboration with researchers from MIT and Harvard, built an AI/ML engine capable of predicting acute deterioration risk up to 24 hours before symptoms would otherwise prompt action, a refined algorithm approaching 99% accuracy, built on the foundation of an earlier clinical study with over 5,000 ICU patients in Singapore. My job was to solve what came next: designing the interface that put that intelligence into the hands of the people who could act on it.
The problem I was handed
The predictive engine was powerful. The interface presenting its output was not. Patient status, SOFA scores, SIRS flags, vitals, risk tables, and clinical graphics competed for equal attention on the same screen. For a medical professional making a care decision at 3am, that cognitive load wasn't an inconvenience, but a clinical risk in itself.
Every unnecessary label, every ambiguous data point, every extra tap was a moment of friction between a clinician and the action they needed to take. The interface had to get out of the way.
What I did
I was brought in with a clear brief: the AI engine existed and worked, the interface did not, and was given design ownership of the mobile interface from the start. Working alongside the design lead and development team at LKMX, I redesigned the key UI components and created the detailed mockups that translated the algorithm's output into something a time-pressured clinician could use without friction: on a mobile device, at any hour, anywhere in the hospital.
• The core design decision was about information hierarchy. I restructured the interface around clinical priority rather than data category; so the highest-risk signals surfaced immediately, and the number of taps required to act on them dropped. What the clinician needed to decide right now became the first thing they saw.
• The unexpected challenge was my own knowledge gap. I had no medical vocabulary when I started. Learning what a SOFA score meant (not just as a label, but as a clinical signal that meant something different at 2am than at 2pm) shaped every design decision I made. Domain immersion wasn't background research. It was the design work.
Design Process
Key wireframes
Key views
Final mockup
The Outcome
The numbers were validated with clinical rigor. The platform's AI engine was built on the foundation of a clinical study involving over 5,000 ICU patients in Singapore, subsequently refined to a 24-hour prediction window with algorithm accuracy approaching 99%. The application received regulatory approval and ISO certification, establishing it as a clinically validated integration of machine learning in patient care.
✔ 30% reduction in patient Length-of-Stay. Average LOS fell by 30% during the 9-month intervention period. When the application was phased out, LOS returned to pre-intervention levels. That controlled reversal confirmed the design's causal contribution: not correlation, not coincidence.
✔ 24-hour early warning anywhere in the hospital. The mobile interface gave healthcare professionals round-the-clock access to predictive risk scores, enabling care escalation a full day before symptoms would have otherwise prompted action, and making critical patient data accessible beyond nurse stations, wherever clinical judgment was needed.
What's next
Peach IntelliHealth was built in 2017, before AI tools were publicly accessible, which makes the clinical outcomes even more striking in retrospect. The platform was ahead of its time, but COVID interrupted conversations with angel investors before the product could scale. In 2026, the original software engineer and I are in active talks to reboot the project, adapting the interface for current AI capabilities, expanding device integration to include Apple Watch, and making the platform more accessible to a broader clinical audience.
What I carried forward
This project permanently changed how I think about information hierarchy. Before Peach, I thought about it as an aesthetic and usability concern. After Peach, I think about it as a decision-support problem: 'what does this person need to decide right now, and is that the first thing they see?' I've applied that question to every significant project since.
If I were doing this project today, I'd run the interface through actual clinical simulation testing rather than standard usability testing: the cognitive load a clinician experiences at 3am managing multiple deteriorating patients simultaneously is different from a usability participant in a controlled remote session. That gap between simulated conditions and real conditions is always where the most important friction hides.
• Tools: Sketch, OmniGraffle (this project predates my current Figma workflow).
• Role: Digital Designer: UI component redesign and detailed mockup creation, in collaboration with the design lead and development team at LKMX.
• Context: Singaporean hospital environment, AI/ML clinical platform. Design work spanned several weeks between 2017 and 2018; 9-month clinical validation period followed launch.
Link: AI for better patient care
