AI Modules

  • Role - Product Designer
  • Project Type - AI-based features for health-based products
  • Status - Ongoing
  • Team - Full-stack Developers, Machine Learning Scientists, Product Managers, External Stakeholders

AI Modules delivers AI-powered experiences for data collection, insights, engagement and model improvement. These modules come in a form of a software development kit (SDK) or a set of APIs that can be integrated into mobile apps. The following are the designs for three of these modules:

Smart Subgroups

Conversational AI for complex, high-stakes data exploration

Smart Subgroups is an AI-powered exploration layer embedded in Aetion's real-world evidence platform, built for researchers and scientists who need to interrogate large, complex patient cohort datasets without building manual queries from scratch. It pairs a conversational AI interface with a live heatmap visualization, letting users ask natural-language questions about their data and receive structured, data-grounded answers in context, alongside the source data that supports them.

Product highlights:

  • Designed the full conversational AI experience: prompt scaffolding, confidence signaling, error recovery, and inline feedback collection
  • Integrated chat responses alongside live data visualizations to support simultaneous reading of AI-generated insights and underlying data
  • Built feedback mechanisms directly into the response flow to support active model improvement from researcher input
  • Designed for high-stakes, regulated contexts where response clarity and appropriate guardrails were non-negotiable requirements

Potential applications include:

  • Exploration - accelerating the path from open research question to structured data insight
  • Accessibility - translating dense subgroup statistics into plain-language summaries usable across technical and non-technical stakeholders
  • Decision support - surfacing condition prevalence differences across subgroups to inform clinical and operational planning
  • Model improvement - inline rating and comment flows that close the feedback loop between researcher input and AI refinement

Pupill

AI for medication data collection

Pupill is a seamless way to gather high quality, coded medication data from users who have their pill bottles or medication packages on hand. It leverages Optical Character Recognition (OCR) and Natural Language Processing (NLP) to output medications, which are user-verified and corrected, then fed through partner medication data pipelines to code and store.

Product highlights:

  • Taking a picture to auto-load medications is more precise and easier than manual entry
  • Prediction-then-verification motion automatically trains/improves the model
  • Dependent on proven OCR/NLP technologies from Google and Amazon

Potential applications include:

  • Safety - identifying drug-drug interactions
  • Substitutions - lower cost options for members, and identifying generics
  • Organizing - accurate and timely medication lists
  • Care - integration into disease and care management programs

Smart Claims

AI for streamlining claim submissions

Smart claims is a computer vision model that leverages photos or digital versions of itemized medical bills to audit whether critical information is present for successfully billing a claim and digitizing that information for downstream use.

Product highlights:

  • Taking a picture and automatically and know instantly if your bill can be processed without being rejected for missing information
  • Reducing rejection rates as far upstream as possible, reducing downstream costs
  • Reducing user burden and errors
  • Making an otherwise stressful and unpredictable experience easy and deterministic for users, and increasing satisfaction

Potential applications include:

  • Out-of-network and international claims submissions

Smart Selfie

AI for basic health information

Smart Selfie acquires basic health information using a computer vision AI model. It predicts age, sex, weight, and height from selfie photos. Associated downstream predictions can include blood pressure, LDL cholesterol, as well as other health predictions associated with standard health predictors.

Product highlights:

  • Taking a selfie and receiving predictions can be an engaging way to gather health information
  • Prediction-then-verification motion automatically trains/improves the model
  • Selfie can also be used for other purposes with user consent (e.g. biomarker training)

Potential applications include:

  • Gateway to myriad health predictions, and acquisition of more health data
  • Onboarding a user to a health related application

Project Takeaways

Acquiring feedback is key to improving the services  of the product and how we can better tailor to and evolve with user’s needs.

‍Acquiring feedback from users is a significant AI pattern seen across many major mobile applications, and this isn't different in the health context. It allows the companies to learn from the data collected and make necessary changes to their platform to improve user experience and ultimately achieve their bottom line (e.g. profit based on increased and consistent engagement). This usually suggests that user feedback is a significant driving force for what the company offers and services, and how it evolves through time. A successful AI design pattern needs to allow for the evolving needs and challenges of users. Users after all dynamic, and often their behaviors are a result of their environment and culture they are surrounded with. For example, Facebook has ways for users to personalize and curate their newsfeed through reporting what they find offensive and the ability to see less of a particular kind of ad or someone’s off-tune memes.

We should expect outputs to be dynamic and hold ourselves accountable for outputs that are incorrect, misleading, or off-base.
‍
Another very common AI pattern is displaying suggestions and recommendations that is catered to the user based on the feedback and data the company has already acquired from them. Reasons for certain suggestions vary. Some are based on who you know (Linkedin), activity and interests (Facebook), history of purchases (Amazon), and geography (Airbnb).Not all search results are based on closely observed interests and history. In the case of navigation apps, search results are often based on a few parameters (e.g. preferred mode of travel, desire to avoid traffic, arrival time, etc.), and the results are live suggestions of optimal routes based on those parameters (Google Maps). While you’re driving, the route can change depending on live input from satellites identifying traffic and other hazards ahead. These solutions aren’t perfect given the dynamic nature of the technology. Sometimes the directions lag, and sometimes they’re plainly just a more inefficient way to get somewhere (tried and true knowledge of backroad shortcuts of how to get from point A to point B is something that doesn’t often get accounted for in navigation).