Bridging survival analysis and machine learning to improve healthy life expectancy estimation using PHR records

Area of Science:

• Public Health
• Biostatistics
• Health Informatics

Background:

• Healthy Life Expectancy (HLE) is crucial for individual quality of life and national healthcare budgets.
• Minimizing time spent in ill-health is a key public health objective.
• Personal Health Records (PHR), integrating Electronic Health Records with lifestyle data, offer a rich resource for health analysis.

Purpose of the Study:

• To investigate the loss of healthy life using Personal Health Records (PHR).
• To directly estimate Healthy Life Expectancy (HLE) using Survival Analysis (SA).
• To develop a Machine Learning (ML) model for predicting loss of healthy life and understand influencing lifestyle factors.

Main Methods:

• Utilized Survival Analysis (SA) for direct estimation of HLE.
• Developed a multiple imputation ensemble Machine Learning (ML) model to predict loss of healthy life within one year.
• Employed ML explainability techniques to identify key lifestyle predictors of healthy life maintenance.

Main Results:

• The ML model achieved an AUPRC nearly double that of random chance on unbalanced data for predicting loss of healthy life.
• ML explainability provided insights into the relationships between lifestyle factors and maintaining healthy life.
• A novel method combining ML predictions with SA hazards was proposed for a conditioned Survival Function.

Conclusions:

• Personal Health Records combined with advanced analytical methods (SA and ML) can effectively investigate and predict loss of healthy life.
• Machine learning models, particularly with explainability, offer valuable insights into modifiable lifestyle factors impacting HLE.
• The proposed hybrid approach allows for tailored survival function estimation, potentially aiding personalized health strategies.