Supervised mixture of experts models for population health

Area of Science:

• Health Informatics
• Machine Learning
• Public Health

Background:

• Observational healthcare data offers a rich source for improving public health outcomes.
• Identifying patient subpopulations and their specific risk factors is crucial for targeted interventions.
• Current methods may not fully leverage the complexity of healthcare data for subpopulation analysis.

Purpose of the Study:

• To develop and evaluate machine learning models for identifying patient subpopulations and their risk factors.
• To improve public health outcomes by deriving actionable insights from observational healthcare data.
• To compare the performance of novel supervised mixture of experts models against existing machine learning techniques.

Main Methods:

• Developed two supervised mixture of experts models: Supervised Gaussian Mixture model (SGMM) for general features and Supervised Bernoulli Mixture model (SBMM) for binary features.
• Applied the models to analyze high cost drivers of Medicaid expenditures for inpatient stays in New York State (NYS) in 2016.
• Compared model performance and insights with state-of-the-art methods including random forests, boosting, and neural networks.

Main Results:

• The proposed SGMM and SBMM achieved comparable prediction performance to existing methods.
• The models successfully identified distinct patient subpopulations and their associated risk factors.
• The SBMM demonstrated superior or equivalent performance for binary features, providing insightful explanations.

Conclusions:

• Machine learning-driven approaches, particularly supervised mixture of experts models, can effectively derive population health insights from observational healthcare data.
• These methods offer a powerful tool for simultaneously identifying at-risk subpopulations and their specific risk factors.
• The findings highlight the potential for improving public health outcomes through advanced data analysis of electronic health records.