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

Social isolation affects one in four people and is associated with adverse health outcomes, yet accurate prediction models remain lacking. This study develops and validates an interpretable machine learning (ML) approach to predict social isolation and identify key predictors among middle-aged and older adults in China.

Training data came from the China Health and Retirement Longitudinal Study. Baseline data from 2011 included 12,785, 12,323, and 11,590 participants for the 2-, 4-, and 7-year prediction models. External validation used the China Family Panel Studies 2010, 2012, and 2016. Five ML algorithms were used to construct prediction models with 283 candidate predictors. SHapley Additive exPlanations explained the feature importance. Classic logistic regression and restricted cubic spline (RCS) explored potential causal associations.

In the development phase, the gradient boosting machine (GBM) performed best across 2-, 4-, and 7-year models (area under the receiver operating characteristic curve [AUC-ROC] = 0.767, 0.729, and 0.749). In the external validation, the GBM had AUC-ROC with 0.649 and 0.678 for the 2- and 7-year prediction models. Age, monthly nonfood consumption, and net primary residence value were consistently identified as the top predictors. Environmental exposures (greenness exposure, rainy days) and community environment (convenience stores, out-migrants) also emerged as important predictors. RCS analysis revealed nonlinear associations between these external factors and social isolation.

With multimodal data, the GBM outperformed existing models for identifying social isolation risk. Its interpretability highlights actionable and potentially reversible targets, especially at community and environmental levels.