Depressive symptoms represent a significant public health concern, particularly among young adults, with increasing prevalence observed globally. While genetic factors play a role, environmental influences are also crucial. Land use, reflecting the urban living environment, is increasingly recognized as a potential factor influencing mental health, but assessing its complex relationship remains challenging. Previous studies employing single indices to quantify the relationship between land use and mental health have yielded inconsistent results, suggesting a more nuanced approach is needed to capture the interaction between different land use types. This study hypothesizes a complex relationship between various land use features and depressive symptoms, unquantifiable by conventional indices. The objectives were to cluster participants with similar urban land use patterns, assess both linear and non-linear relationships between land use and depressive symptoms, and identify potential differences in these relationships between clusters. The study leveraged the FinnTwin12 cohort, a rich dataset offering longitudinal information on Finnish twins, to address this research gap.

Existing research on the link between urban land use and mental health is characterized by inconsistent findings and methodological limitations. Some studies suggest specific urban environmental profiles, including land use density, influence mental health through brain volume and biological pathways, while others find associations between the urban environment and the incidence of serious mental illnesses. However, inconsistencies remain, largely due to the limitations of conventional indices that struggle to capture interactions between diverse land use types and often fail to account for the complexities of high-dimensionality and small effect sizes inherent in this area of research. The need for sophisticated multi-exposure models capable of handling non-linear relationships is widely acknowledged, especially in light of the success of machine learning methods in studying urban exposures and other health outcomes. Yet, the application of such methods in mental health research remains relatively rare.

This study utilized data from the FinnTwin12 cohort, focusing on 1804 twins living in urban areas of Finland in 2012. Depressive symptoms were measured using the General Behavior Inventory (GBI) in young adulthood (mean age 24.1). Land use data was derived from Urban Atlas 2012, encompassing eight land use types within three buffer radii (100m, 300m, and 500m) around each twin's residence. A land use mix index was also calculated. Seven demographic covariates (sex, zygosity, parental education, smoking, work status, secondary school level, age) were included, along with four social indicators (age structure, education level, unemployment rate, income level) at the postal code level. K-means clustering was employed to group participants based on their land use profiles, resulting in two clusters. The sample was then split into training and testing subsets for model building and validation. Linear elastic net penalized regression was used to assess linear associations and identify important land use features, adjusting for both demographic and social indicators. eXtreme Gradient Boosting (XGBoost) was employed to explore non-linear relationships, and SHAP values were used for model interpretability. A sensitivity analysis using linear mixed models controlled for genetic effects by treating twin pairs as random effects. Finally, a post-hoc linear regression analyzed the association between the land use mix index and depressive symptoms.

K-means clustering identified two distinct clusters: one characterized by a higher proportion of high-density residential land use (Cluster 2, city center-like) and another with more low-density residential land use (Cluster 1, suburban-like). Linear elastic net regression revealed heterogeneous patterns across the overall sample and the two clusters. In Cluster 1 (suburban), agricultural residential land use within a 100m buffer was most strongly associated with higher GBI scores after adjusting for demographic covariates. After further adjustment for social indicators, additional land use features (urban green, natural areas, high-density residential and water areas) showed associations. However, in Cluster 2 (city center), no land use exposures were significantly associated with GBI after either adjustment. Refitting to a linear mixed model did not significantly alter the results in Cluster 1, except it showed commercial and industrial land use in a 300m buffer to be positively associated with GBI before further adjustment. XGBoost models confirmed the heterogeneity, showing natural land use in a 100m buffer (Cluster 1) and infrastructure land use in a 300m buffer (Cluster 2) as the most influential factors after further adjustment. SHAP value plots revealed non-linear relationships between land use exposures and depressive symptoms. Post-hoc analysis using the land use mix index showed a significant association with depressive symptoms in Cluster 1 before adjustment, but this association disappeared after adjustment. Model performance (RMSE) was good across both models and clusters.

This study's findings reveal a complex and context-dependent relationship between urban land use and depressive symptoms in young adults. The lack of association in the city center cluster (Cluster 2) may be explained by differential access to healthcare, social support systems, or other environmental factors. The positive association of agricultural residential land use with depressive symptoms in the suburban cluster (Cluster 1) could be linked to factors like longer commutes and potential isolation. However, this requires further investigation. The importance of considering the social environment is highlighted, as its inclusion strengthened the associations observed in the suburban cluster. The study also underscores the limitations of using single indices, such as the land use mix index, which did not consistently capture the complexities of the land use-mental health relationship. The use of multiple statistical approaches, including machine learning, effectively handled the high-dimensionality and potential for non-linearity and uncovered nuances missed by traditional methods. Future studies should investigate these interactions further within their specific contexts.

This study is the first, to our knowledge, to use a multi-model approach to explore the complex interplay between diverse urban land use features and depressive symptoms in young adults. Findings highlight the importance of considering population heterogeneity and contextual effects. The results emphasize the need to move beyond simplified indices and embrace more complex models and data-driven approaches. Future research could investigate the mediating role of social factors and explore the temporal dynamics of these relationships. Further studies with larger sample sizes and longer follow-up periods are needed to validate these findings and establish causality.

The study has some limitations. The cross-sectional nature of the data prevents causal inferences, and the relatively small sample size, while addressed by the robust methods employed, calls for replication in larger studies. Genetic influences were partially controlled for but not fully disentangled. Confounding effects from other environmental exposures (air pollution, noise) were not fully addressed. The interpretability of the machine learning models, while enhanced by SHAP values, remains a challenge. These limitations highlight the need for future longitudinal studies with comprehensive environmental assessments to confirm and extend these findings.