Machine Learning to Advance Individual and Population Health: A BMC Innovation Story

About the Project

‍

Boston Medical Center (BMC) is committed to empowering all patients to thrive, through innovative and equitable care. With sophisticated, inclusive, and patient-centered research, BMC believes it can improve health outcomes and drive change. It's why the institution wanted to better analyze patient data for enhancing patient care.

‍

Through a collaborative effort with Appsilon, a research team at BMC leveraged our machine learning and life science expertise to enhance their data analysis capabilities with the goal of improving health outcomes and population health. This resulted in the successful second iteration of their project and the development of a robust proof of concept that lays the groundwork for rapid development of new approaches to support decision-making through predictive analytics and improved patient categorization, uncovering new healthcare insights.

‍

‍

Background

1. The BMC research team contacted us to enhance their R/Shiny app, “The Health Equity Explorer (H2E)”, to better support using clinical and “place-based” data for people living in the Boston area, focusing on health equity and outcome disparities, This work aligns perfectly with our Data4Good initiative. Our previous collaboration resulted in an impressive R Shiny app that fulfilled their initial objectives.
2. Following the successful completion of the first iteration, we embarked on a second phase that expanded the system to include a broad range of social and environmental variable related to where patient lived and a new “Advanced Analytics” module that is an easily extensible framework to support machine learning.
3. Some statistical techniques, such as logistic regression, were employed in the earlier project. However, it was evident that there was much more potential to explore in terms of both supervised and unsupervised learning methodologies.

Tech Stack

‍

• Data Formats: Database based on The OMOP (Observational Medical Outcomes Partnership) Common Data Model
• Python
• Jupyter Notebooks

‍

‍

Meeting BMC’s Needs 

‍

BMC's primary objective was to leverage existing work using the OMOP common data model to expand the system to be able to support multiple common data models. In this way, a broad range of academic medical sites across the US and internationally could use the open-source platform. This approach allowed for seamless data integration and manipulation, aligning with their goal of examining health outcomes over time. They sought to break down these outcomes into informative segments, enabling more informed decision-making regarding model inclusions.

‍

Their specific needs could be distilled into three key aspects:

‍

• Health Outcomes: The ability to adapt patient grouping strategies.
• Advanced Analytics: Access to predictive probabilities and odds’ ratios for selected variables, addressing a limitation they faced with Tableau.
• GIS/Place-based Analysis: Incorporating maps to study health outcomes by neighborhood and gaining deeper insights into neighborhood-specific metrics.

‍

‍

Key Initiatives

‍

Our engagement with BMC encompassed three primary areas of interest:

• Feature Selection: Allowing users the ability to explore the relative importance of over 1000 demographic, clinical, and place-based variables in an interactive and iterative way via ML
• Predictive Analysis: Exploring individual patient health for tailored treatment approaches.

• Patient Group Clustering: Understanding the broader context by clustering patients to understand their diverse needs.

‍

While BMC conceptualized these objectives, they sought our expertise in translating them into practical solutions, emphasizing the need for explainability.  

‍

We have successfully developed our Feature Selection Module and have a framework that will allow rapid development of new solutions that entail clustering patients into groups based on their information and subsequently building a classifier to elucidate the defining characteristics of each group. This approach offered a population-level view with individual-level explanations, achieved through visualizing decision trees and feature importance.

‍

‍

Plan of Action

‍

To ensure the success of this study, we established a clear goal: demonstrating the application of supervised and unsupervised machine learning methods on patient data while prioritizing explainability. 

‍

We aim to create a flexible framework that accommodates diverse datasets, allowing users to select features and encoding methods for population clustering and classification. User-friendliness and adaptability to various data types were paramount considerations.

‍

Our data analysis process consisted of:

• Data Visualization using PCA: Employing Principal Component Analysis (PCA) to provide an initial understanding of data structure.
• Clustering Algorithms: Integrating K-Means, K-Modes, and K-Prototypes to cater to the heterogeneous nature of medical datasets.
• Decision Tree Classifier: Training a decision tree classifier on the clustered patient groups to understand feature importance and provide explanations.
• Model Interpretability: Leveraging the SHAP package to explain feature importance between clusters and individual patient-level contributions to model predictions.

‍

This project addressed BMC's immediate needs and paved the way for more informed decision-making and enhanced patient outcomes by applying advanced machine learning techniques.

‍

‍

Challenges and Expectations 

‍

BMC recognized the potential of machine learning for predictive analysis, and sought expertise around the specific methods and techniques to apply for their data environment.

‍

This POC has guided them to valuable predictive analysis opportunities with clinical data, escaping the limitations of alternative software solutions.

‍

‍

Results and Impact

• A novel ML framework and fully functional “Feature Selection” Module.
• Proof-of-Concept machine learning models within a week tailored to individual patient data to improve patient outcomes that can be easily added in future versions
• Improved strategic decision-making through predictive analytics based on comprehensive patient data analysis, leading to optimized resource allocation.
• By integrating social and environmental determinants into predictive models, healthcare professionals can better understand and address the contributions of various underlying factors such as race, age, disability, and social class, leading to improved healthcare equity regardless of socio-economic status and environmental conditions.
• Adopted a common data model, enhancing the scalability and transferability of healthcare analytics, paving the way for broader application and impact in the industry.

‍

‍

Insights from Domain Experts

‍

‍

Challenges and pain points:

‍

We were excited to think about how machine learning techniques and more advanced analytics could be used to automate, explore, and perhaps discover new insights that we weren't able to achieve with the more traditional approaches.  We've demonstrated pretty clearly that the approach is feasible, that the tools can be applied fairly quickly, and that there's a lot of possibility and opportunity.

- Bill Adams, Boston Medical Center

‍

‍

Future Directions

‍

The initial success of the POC has sparked innovative ideas for expanding this work by incorporating additional machine learning features.

‍

This also demonstrates a simplified application of supervised and unsupervised learning techniques. Numerous avenues exist for extension and further exploration, enabling a deeper understanding of patient data.

‍

‍

‍