A machine learning approach to identify the gender-specific risk and protective factors of loneliness among adults in Shanghai, China

Loneliness is increasingly acknowledged as a global health priority and a significant policy concern. This study aimed to identify and rank the gender-specific risk and protective factors for loneliness among Chinese adults. We employed a combination of machine learning (ML) and traditional statistical approaches to compare 34 potential risk and protective factors for loneliness among 3220 adults aged 18+ (mean age 34.4 [SD 10.9]; 48.5% women) in Shanghai, China during 2022. Loneliness was measured using the UCLA Loneliness Scale (scores ≥ 7 indicating loneliness). Predictors spanned demographic, employment, financial, psychosocial, lifestyle, and health domains. We compared multiple ML algorithms (support vector machines, gradient boosting, Naïve Bayes, Random Forest, and K-Nearest Neighbors [KNN]) and selected a KNN classifier for its superior performance. Using Shapley Additive Explanations (SHAP) values from the KNN model, we identified the most important predictors of loneliness, followed by stepwise logistic regression to quantify the explained variance and calculate odds ratios (ORs) for the most significant predictors. Poor sleep quality, self-rated poor health, and low social capital emerged as the strongest risk factors, collectively explaining 19.8% of the loneliness variability among women and 17.1% among men. In women, poor sleep quality (OR 1.61; 95% CI 1.30–2.00), self-rated poor health (OR 1.39; 95% CI 1.15–1.69), and low social capital (OR 1.35; 95% CI 1.13–1.61) were significantly associated with loneliness. Similar patterns were observed in men. Gender-specific factors included less frequent social contact in women and intimate partner violence in men. COVID-19-related income loss significantly increased loneliness risk, particularly in men. Among 34 predictors, poor sleep quality, self-rated poor health, and low social capital were robust risk factors for loneliness across genders. These results underscore the importance of gender-specific approaches to address this growing health concern.