testimonialManifold’s contribution has been invaluable. Our integration of machine learning algorithms will enable the dynamic adjustment of treatment, and we are able to better deploy human support at the right time and in the right way.

— Mark Berman, MD FACLM, Head of Health


Better Therapeutics uses software to treat and reverse diabetes, hypertension, and other metabolic diseases by changing behaviors at the root cause. Their solution integrates a mobile medical app that delivers highly personalized behavioral therapy augmented by health coaches, with clinical decision support software for physicians.

They engaged Manifold to help solve two problems: first to empower coaches with data so they could effectively prioritize patients, and then to provide data-driven insights on which interventions to recommend to targeted patients. Patient data was too complex for coaches to effectively interpret on their own; they needed AI to quickly process the data and deliver the right recommendations.

Manifold accelerated development of HIPAA-compliant, AI-powered features to deliver customized treatment plans for patients. The product consists of an AI engine embedded in an application used by coaches when working with patients, which saves them time and effort in determining the optimal therapy strategy for a given patient.


Manifold worked closely with the client’s product and medical teams to establish which metrics mattered for a patient’s health issue and what the target changes in these metrics should be. We used two totems of health for diabetes: the patient’s weight, and a glucose biomarker called A1C. Manifold developed a pair of models to predict the probability of hitting key desired outcomes after 16 weeks in both categories for each patient under treatment.

Feature Identification

In order to help make this prediction, we needed to understand what features drive outcomes. We worked with the Head of Health to identify predictive features. We hypothesized the key features would be engagement with the healthcare app. With this hypothesis in mind, we created features that were proxies for engagement.

Model Engineering

Manifold built a pair of models using these features, starting with a random forest and moving to a gradient boosting tree. We performed this work using the Python ML stack of Pandas, Scikit-learn, and other open source data science tools. We then worked with the client to build and deploy our solution inside Datica, a HIPAA-compliant cloud they were already using. We built a REST API using Flask to achieve production-grade machine learning, allowing the client to run scheduled jobs.

Better Therapeutics had hundreds of patients in a clinical trial, which gave us the benefit of working with ground-truth data. Despite an initially modest-sized data set, we were able to perform a sound analysis using five-fold cross-validation. By weeks 4–8, the algorithm yields an AUC in the .8 to .9 range.

BI Tool Support

After building the AI predictive models, Manifold worked to support the client's development team on delivering those predictions to coaches and patients. We offered guidance as they used Looker, a BI data visualization creator, to make a tool that helps their health coaches.


Manifold augmented the scale and abilities of our client’s coaches to deliver personalized patient care by helping to predict which patients would need additional support and intervention to meet their goals. Coaches can now put their focus where it’s needed most, and are better able to ensure consistency of care with on-demand access to the recommendation engine. And patients benefit from coaches who can now spend more time listening to their specific needs and concerns.

Better Therapeutics found our work to be both valuable and reusable: they have since made the prediction metric available back to the patient as a “health score” so that the patient can track their own progress with a single number, like a credit score. This project also led to downstream applications, with multiple integration points for the predictions from our models. Indeed, they continue to engage with Manifold to build out additional machine-learning enabled features.

Key Takeaways

  • Worked to understand predictive features and create proxies for engagement

  • Built a REST API using Flask, allowing the client to run scheduled jobs

  • Augmented the scale and abilities of coaches to deliver personalized patient care