Algorithmic-Powered Health Products
Algorithmic models can appear robust in development but have limitations in new user contexts. This delivers inconsistent and less relevant outputs across populations.
When training data doesn’t reflect real-world populations, models can show significant drops in performance to influence:
· Reduced user and clinical trust
· Lower adoption and engagement
· Increased risk of bias and regulatory scrutiny
I design data curation and training strategies that address these gaps, improving performance and acceptance across populations and enabling more equitable deployment.
Use Case 1: Data Curation to Train Human-Centered AI
I lead the design of a data curation and training strategy that uses human-centered taxonomies and psychosocial and cultural theories to better reflect real-world populations. By planning what data is included, how it is labeled, and which segments must be visible in the training process, the work reduces risks like perpetuating stigma, delivering inaccurate or unsafe advice, and missing context that matters to users.
This approach treats AI model training as a design decision, not a technical afterthought. It improves the relevance and equity of outputs across populations and supporting safer, more accountable model deployment.
Use Case 2: Health AI That People Use and Trust
I apply my work in eHealth literacy to understand how people with different knowledge, skills, and abilities engage with AI-supported tools. This includes clarifying what users need to know to interpret outputs, what skills they require to act on guidance, and how this translates to offline behavior.
It turns AI from a clever demo into a dependable product that is useful for humans. Further, it allows people to understand and act on outputs, trust the system appropriately, and harmful or reputation‑damaging edge cases are less likely to improve engagement, task completion.