adhereQI by Domlytics

AI-powered tool predicting medication adherence using social determinants data—boosting patient outcomes, reducing costs, and transforming proactive healthcare decisions.

Comment

Inspiration

I have worked in Quality Improvement Data Analytics for most of my career. I am passionate about Digital Quality and Health Equity. When this email hit my inbox, it did so at an interesting time. I had been debating leaving my senior executive role in clinical analytics for a health plan and starting my own business. That is when I decided I was going to do it, and this app would be my first product I would bring to market, and it is only a single module of a much larger solution I intend to build for smaller regional plans and provider groups/offices who don't have the analytical teams to support the same initiatives the bigger organizations do. I want to make Data Science and AI in healthcare attainable to the small to mid size organizations, practices, and systems and ensure that as many patients as possible can have improved outcomes.

What it does

adhereQI uses an advanced machine learning algorithm to predict medication adherence risks by analyzing thousands of data points from publicly available social determinants of health (SDOH) databases. The model looks at data points such as income levels, insurance coverage, transportation access, education, and healthcare availability among many others. It delivers actionable insights such as a Medication Adherence Risk Stratification (High, Medium Low), Medication Adherence Probability Score, Pharmacy Access Score, and Key Social Determinants for the patient.

How we built it

Because the training data would be considered PHI and I did not have enough time to obtain Data Use Agreements from CMS for deidentified data, I used over 114,000 FHIR based synthetic patient data using the open source Synthea package link. I also sourced 10 years of AHRQ SDOH indicators. These were loaded into an Azure PostgreSQL database for our model to reference during training and makes it easily updated for retraining which I plan to do for continuous improvement and continuous delivery.

For Machine learning I leveraged XGBoost and Optuna to uild and optimize a predictive model with robust performance and minimal overfitting, I plan on retraining this model with deidentified patient data to further improve performance and introduce real world scenarios that may have not be captured by the synthetic dataset.

I also included an interactive dashboard built with Flask to provide intuitive visualizations to help facilitate clinical decison making.

Challenges we ran into

Data quality and integration was probably my biggest obstacle. Ensuring that joins were accurate and not degrading referential integrity or data quality from disparate data sets was challenging, and also working with API calls is not something I typically have had to do in my career as I have usually had data already in databases or in a data warehouse to work with.

Another major obstacle was carefully reviewing features for my model. My first models were achieving suspiciously high accuracy, many cases 100% accuracy. This required heavy feature engineering and tuning and some exploratory analytics to cross-validate my stratigies with observations of the underlying data.

I also had some GPU/CPU conflicts with my model. I am running the model on a PC I built running on Ubuntu LTS and running an AMD Ryzen 9 7950x3d CPU, 128GB DDR5 RAM and an RTX 3060 GPU. The models thankfully were still very performant when running on the CPU thanks to my hardware, but it did require some parameter adjustments to get the model running efficiently.

Accomplishments that we're proud of

I am very proud of having developed a fully functional and visually engaging web application for my minimum viable product and i am excited to continue to deliver improvements as time goes on.

What we learned

I definitely got some hands on experience with hyperparameter tuning for complex machine learning models and how to develop a full stack end to end web application using different environments, languages, and deploying it with docker.

What's next for adhereQI by Domlytics

In addition to continuing to improve the Medication Adherence model, I have planned to add a patient readmission risk model, and eventually build out a CMS Stars web application that can support not only clinical decision making but also strategic and informed decision making for system administrators and health plans to improve their CMS Star ratings and outcomes and improve their health equity index as well. I also would like to incorporate FHIR Davinci CDex as well to help improve data interoperability and collaboration between clinicians, payers, and systems.

Share this project:

Updates