Predicting onset of diabetes with a Healthtech innovator

96 million U.S. adults are on the brink of developing diabetes, and 1 in 3 already have it, yet conventional solutions like continuous glucose monitors are reactive tools built for managing the disease, not preventing it. The client set out to close that gap with a solution that could predict onset up to 5 years in advance, which meant building and continuously fine-tuning an AI/ML model for both accuracy and product-readiness, and timeboxing every milestone to stay on schedule.

Our product, design, and engineering pods actively collaborated with stakeholders to launch the Minimum Viable Product (MVP) version. Our partnership with this healthtech innovator was split into two phases:

• Phase 1 - proof of concept: The initial phase of our partnership focused on validating the idea and creating a concrete proof of concept to demonstrate its feasibility.
  A.Customized design sprint: We kicked off the partnership with a customized design sprint inspired by the Google Ventures Design Sprint to dig into the challenges faced by patients, healthcare providers, and insurance companies.
   B. Diverge and converge:  Applying the double-diamond approach and deep research, we created space for brainstorming and prioritization, converging to focus on primary and immediate pain points.
   C. Golden path: We used the Jobs-to-be-Done (JTBD) framework to map out the golden path, followed by building, testing, and iterating the proof of concept via a prototype with pilot customers.

• Phase 2 - MVP to scale: We had some early wins with positive response from early adopters with the prototype. We then put the ML algorithm into production and created a solution that could predict the onset of diabetes in patients with near-perfect accuracy, like no other solution in the market. The learning and improving the accuracy of the model was of course a priority. Leveraging mature engineering practices, this is how we partnered with our  customer to expand product capabilities:
  ‍ A. Data sourcing: We aggregated data from OpenEHR and a partner hospital. Of course all the data was obfuscated to maintain compliance with HIPAA standards of privacy. We cleaned it up and filled in the gaps in the data using imputation techniques.
  B. Feature selection and engineering: We worked with the client to identify high-impact features from a wide range of potential predictors and built them for improved performance.
  C. Algorithm development: Our engineers built a fast-learning, self-improving algorithm utilizing the Accelerated Failure Time (AFT) survival regression model due to its suitability for modeling individual time-to-event data. This choice proved instrumental in achieving 85% accuracy in predicting the onset of diabetes.
  D. Multi-modal data fusion: We utilized multiple modes of biometric and demographic data from various sources, including hospital visits and social history, followed by streamlining, processing, and feeding data into the model to facilitate continuous training.
  E. Continuous Improvement: With each new iteration and implementation, we ensured that the algorithm continued to learn and improve its predictions, resulting in a continuously evolving and highly accurate solution.
  F. Robust CI/CD pipelines: A Continuous Integration/Continuous Deployment (CI/CD) pipeline was deployed to shorten feedback loops and consistently meet product engineering milestones.
  G. Scalability: We designed the solution to be easily scalable by streamlining data ingestion and processing, ensuring that it could handle increasing data volumes without compromising performance.

Zemoso has always been at the forefront of AI innovation across industries, helping innovators solve critical problems by leveraging the power of data, AI, and ML - and this partnership is a clear example of that in healthcare.