Development of a Predictive Model for the Onset of Loneliness in Older Adults From the National Center for Geriatrics and Gerontology-Study of Geriatric Syndromes.
Loneliness is a rising public health concern among older Japanese adults. We aimed to construct a predictive model for loneliness onset among older adults and evaluate its predictive performance.
A total of 4050 participants responded to our survey (mean follow-up period: 3.1 [range, 2.8-3.3] years). Of these, 1806 older adults (age ≥ 65 years) who were not lonely at baseline were included. Loneliness was assessed using the UCLA Loneliness Scale (Version 3). A score of ≥ 44 indicated the presence of loneliness at follow-up. Predictive models for the onset of loneliness were developed using 12 machine-learning algorithms. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs). Sensitivity analyses examined the contribution of baseline UCLA-LS information, model stability, and class imbalance handling.
In total, 421 respondents (23.3%) reported loneliness at follow-up. The XGBoost model achieved the highest AUC of 0.740 (0.705-0.776), with an accuracy of 0.631, sensitivity of 0.806, and specificity of 0.577. SHAP analysis indicated that baseline UCLA-LS items were among the most influential predictors. When all baseline UCLA-LS items were excluded, the test AUC decreased to 0.635. Ten-fold cross-validation showed a mean AUC of 0.742 ± 0.065 for the full XGBoost model.
XGBoost showed moderate discrimination for predicting 3-year loneliness onset. However, baseline UCLA-LS information contributed substantially to model performance, and external predictors alone showed modest discriminative ability. Further refinement and external validation are needed before implementation.
A total of 4050 participants responded to our survey (mean follow-up period: 3.1 [range, 2.8-3.3] years). Of these, 1806 older adults (age ≥ 65 years) who were not lonely at baseline were included. Loneliness was assessed using the UCLA Loneliness Scale (Version 3). A score of ≥ 44 indicated the presence of loneliness at follow-up. Predictive models for the onset of loneliness were developed using 12 machine-learning algorithms. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs). Sensitivity analyses examined the contribution of baseline UCLA-LS information, model stability, and class imbalance handling.
In total, 421 respondents (23.3%) reported loneliness at follow-up. The XGBoost model achieved the highest AUC of 0.740 (0.705-0.776), with an accuracy of 0.631, sensitivity of 0.806, and specificity of 0.577. SHAP analysis indicated that baseline UCLA-LS items were among the most influential predictors. When all baseline UCLA-LS items were excluded, the test AUC decreased to 0.635. Ten-fold cross-validation showed a mean AUC of 0.742 ± 0.065 for the full XGBoost model.
XGBoost showed moderate discrimination for predicting 3-year loneliness onset. However, baseline UCLA-LS information contributed substantially to model performance, and external predictors alone showed modest discriminative ability. Further refinement and external validation are needed before implementation.
Authors
Shimoda Shimoda, Katayama Katayama, Yamaguchi Yamaguchi, Nakajima Nakajima, Kawakami Kawakami, Yamagiwa Yamagiwa, Akaida Akaida, Shimada Shimada
View on Pubmed