Predicting and explaining social isolation: insights from an interpretable machine learning model in ageing...

Area of Science:

• Gerontology
• Public Health
• Artificial Intelligence

Background:

• Social isolation affects a quarter of adults and is linked to poor health outcomes.
• Existing risk prediction models for social isolation are insufficient.
• Developing accurate and interpretable models is crucial for identifying at-risk individuals.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) approach for predicting social isolation.
• To identify key predictors of social isolation in middle-aged and older adults in China.
• To explore potential causal associations between predictors and social isolation.

Main Methods:

• Utilized data from the China Health and Retirement Longitudinal Study (CHARLS) for model development.
• Employed five ML algorithms, including Gradient Boosting Machine (GBM), with 283 candidate predictors.
• Applied SHapley Additive exPlanations (SHAP) for feature importance and restricted cubic splines (RCS) for causal association exploration.

Main Results:

• The GBM model demonstrated strong performance in predicting social isolation risk (AUC-ROC up to 0.767 in development, 0.678 in external validation).
• Consistent top predictors included age, monthly non-food consumption, and net primary residence value.
• Environmental factors like greenness exposure and community characteristics also emerged as significant predictors.

Conclusions:

• The developed interpretable ML model, particularly GBM, shows superior performance in identifying social isolation risk compared to existing methods.
• The model's interpretability highlights actionable and potentially reversible targets for intervention.
• Community and environmental-level factors are crucial targets for mitigating social isolation.