ARDA
AI-Powered Eye Screening
A Problem Worth Solving
Thailand has a rapidly escalating diabetes burden, with prevalence rates surging sixfold in the last two decades. Diabetic retinopathy (DR) is a significant cause of vision loss in Thailand, with an estimated 500,000 people at risk of blindness, particularly in rural areas. Across Thailand, there’s a single ophthalmologist for every 3,000 patients with diabetes. This means retinal screenings, the difference between early treatment and preventable blindness, are being delayed by weeks, sometimes months due to provider limitations. At Google Health, we developed Automated Retinal Disease Assessment (ARDA), a high-accuracy deep learning model that detects diabetic retinopathy at specialist level and we designed the patient experience and clinical implementation in partnership with clinics across Thailand to speed time from screening to treatment.
Making the Bet
As the UX lead for Imaging & Diagnostics, I made the strategic bet to allocate people and budget toward fieldwork in Thailand, recognizing that the model’s success depends on human-centered deployment, not just algorithmic accuracy. I hired and mentored a multidisciplinary squad of researchers and designers, empowering them to conduct research with 11 Thai clinics
Developing alliances
Implementing ARDA required navigating the perspectives and needs of US-based engineering and product teams alongside the clinical staff in Thailand: IT owners, camera technicians, nurses, and doctors. I advocated for a paired research approach, including an engineering lead and product manager clinician in our research planning and execution. As a team we met weekly via video conference with stakeholders in Thailand, developing our collaboration approach and creating a groundwork for in-clinic field research. This collaboration across organizations simplified the product development process by ensuring agreement on service mapping and clinical service design.
Scaling through Systems & Strategy
To ensure ARDA could scale from pilot to platform, I oversaw the creation of a specialized Design System for Imaging & Diagnostics. My goal was to move the team toward a clear-at-a-glance philosophy that prioritized cognitive ease for overextended clinicians.
Environmental Adaptability: I directed the team to stress-test the system for extreme conditions, ensuring high-legibility in both the high-light environments of intake desks and the low-light clinical settings required for retinal imaging.
Global Scalability: By standardizing components and interaction patterns, we ensured the system was localized and culturally adaptable across multiple geographies, from rural Thailand to high-volume clinics in India.
Judgment Calls: Balancing Safety and Adoption
As a leader, my role was to create the strategic space for uncomfortable research findings that challenged our roadmap. I oversaw two pivotal shifts where user data required us to push back against initial product assumptions:
The Interface Pivot: Early prototypes prioritized diagnostic results, treating the fundus image as secondary. My team’s research demonstrated that for nurses, the image was the primary tool for patient counseling. I directed a pivot to treat the retinal image as the hero of the UI, transforming the tool from a data display into a persuasion aid that helped patients with minimal symptoms understand the urgency of specialist care.
Operationalizing Fieldwork into Protocol: When our researchers discovered that nurses were "quietly working around" referral protocols to protect patients from unnecessary travel burdens, I didn't just authorize a UI fix; I drove a clinical protocol change. I aligned our Engineering and Regulatory leads to implement a new overreader review stage for ungradable images, a decision that directly reduced referral burdens while maintaining strict safety standards.
Impact: Scaling Vision Care
By prioritizing human-centered deployment over purely algorithmic benchmarks, we transformed a lab-based model into a life-changing service:
Clinical Efficiency: We successfully shrank the referral pathway from 10 weeks to approximately 10 minutes, enabling same-day clinical decisions in underserved regions.
System Capacity: The project proved that specialist-level AI could be operated effectively by nurses with minimal ophthalmology training, effectively removing the primary bottleneck for patient outcomes.
Academic & Industry Leadership: Beyond the product, I mentored the team through the publication of our findings at ACM-CHI, which has since accumulated nearly 400 citations. This work established a new reference point for human-centered AI deployment in the field.
Future Reach: In 2024, these foundations enabled Google Health to announce partnerships to conduct six million AI-powered screenings over the next decade, focusing on preventing avoidable blindness in Thailand and India.
This project reinforced my experience that the most complex challenge in AI isn't the model's accuracy. The real challenges are trust and integration within human-oriented systems. Leading this team taught me how to scale high-stakes technology by building systems that empower, rather than replace, the people on the front lines.
Selected Works
