Social Isolation Among Economically Inactive Middle-Aged Adults Living Alone In South Korea: A Latent Class and Machine Learning Approach

Background/Purpose: Over the last decade, social isolation has emerged as a criticial social risk in South Korea, driven by the rapid growth of single-person households , an aging popluation, and a rise in economically inactive individuals. Accordingly, a large body of resesarch has explored potential farctors associted with social isoaltion and its conseqeunces mainly among youth and the elderly. However, economically inactive middle-aged adults living alone (EIMALA) remains an under-researched demographic in current social isolation literature, presenting a critical gap in our understanding of middle-aged vulnerability. This group is particulalry vulnerable to social isolation as they are excluded from the primary mechanisms of social integration: workplace-based ties and family-based private support networks. Moreover, most prior studies have examined factors associted with social isolation based on linear regression models; however, they are limited in detecting threshold-based nonlinear patterns. This study identifies latent types of social isolation among EIMALA and examines key factors predicting class membership using a machine learning approach. The fidnings provide empirical evidence to support early intervention for high-risk indivdiuals and suggest the strategic allocation of limited welfare resources.

Methods: Data came from the 2023 Community Health Survey about Korean population. The final analytic sample consisted of 3,063 economically inactive adults aged 40 to 64 living alone. All analyses were weighted to account for the survey’s complex sampling design. Latent class analysis was conducted to classify distinct patterns of social isolation indicators. Social isolation was measured using two dimensions: a three-item social network scale and a four-item social participation scale. To build a predictive model of social isolation class membership, six machine learning algorithms were compared: Gradient Boosting, Random Forest, Extra Trees, LightGBM, XGBoost, and Decision Tree. After selecting the best-performing model, the importance of 20 variables was examined, and SHAP (SHapley Additive exPlanations) values were used to assess the direction and nonlinear patterns of key predictors.

Results: Two latent classes of social isolation were classified: a high-risk group and a low-risk group. The 10 most important predictors of social isolation class membership were identified: in order, monthly income, trust in neighbors in the community, depression, gender, subjective stress, number of walking days, receipt of health screening, age, body mass index, and housing type. SHAP analysis further showed that monthly income and subjective stress exhibited nonlinear relationships with social isolation risk. In particular, the probability of belonging to the high-risk group declined markedly around the monthly income threshold of KRW 1 million (700$). Subjective stress also showed a threshold-like pattern, suggesting that social isolation risk increased sharply at a specific level of perceived stress.

Conclusions/Implications: The findings highlight the heterogeneity of social isolation among EIMALA and identify key predictors that would be useful for targeted interventions. These findings underscore the need for social work practice and policy to prioritize proactive outreach for high-risk cohorts, implement tailored interventions to mitigate social isolation, and optimize the distribution of scarce welfare resources using empirically derived risk factors to social isolation.