Predicting Depression and Anxiety Risk in Dutch Neighborhoods from Street-View Images

TL;DR

This study investigates predicting neighborhood-level depression and anxiety risk in the Netherlands from Google Street View Images using deep learning. By fine-tuning ResNet50 and DeiT Base on 9,879 SVIs linked to Dutch Health Monitor risk metrics, the authors achieve moderate accuracy with substantial adjusted accuracy by ignoring adjacent misclassifications, and employ SHAP and attention rollout to probe interpretability. While landscape cues such as greenery, sky, and built environment influence predictions, the explanations remain ambiguous and not clearly tied to specific risk categories, highlighting interpretability challenges. The work demonstrates the potential of SVIs for environmental monitoring of mental health risk and proposes future directions to improve explainability and enable time-series neighborhood risk tracking across urban and rural settings.

Abstract

Depression and anxiety disorders are prevalent mental health challenges affecting a substantial segment of the global population. In this study, we explored the environmental correlates of these disorders by analyzing street-view images (SVI) of neighborhoods in the Netherlands. Our dataset comprises 9,879 Dutch SVIs sourced from Google Street View, paired with statistical depression and anxiety risk metrics from the Dutch Health Monitor. To tackle this challenge, we refined two existing neural network architectures, DeiT Base and ResNet50. Our goal was to predict neighborhood risk levels, categorized into four tiers from low to high risk, using the raw images. The results showed that DeiT Base and ResNet50 achieved accuracies of 43.43% and 43.63%, respectively. Notably, a significant portion of the errors were between adjacent risk categories, resulting in adjusted accuracies of 83.55% and 80.38%. We also implemented the SHapley Additive exPlanations (SHAP) method on both models and employed gradient rollout on DeiT. Interestingly, while SHAP underscored specific landscape attributes, the correlation between these features and distinct depression risk categories remained unclear. The gradient rollout findings were similarly non-definitive. However, through manual analysis, we identified certain landscape types that were consistently linked with specific risk categories. These findings suggest the potential of these techniques in monitoring the correlation between various landscapes and environmental risk factors for mental health issues. As a future direction, we recommend employing these methods to observe how risk scores from the Dutch Health Monitor shift across neighborhoods over time.

Figures (15)

• Figure 1: Geographical distribution of neighbourhoods represented in the dataset.
• Figure 2: Confusion matrices of the predictions of fully fine-tuned models. a) ResNet50. b) Deit Base.
• Figure 5: Distribution of the target variable. Left: the histogram of the number of neighbourhoods with respect to the depression and anxiety risk scores. The histogram is split into 5 bins in order to represent the distribution of the target variable’s classes. Right: box plot showing the median, the quartiles and the spread of outliers of the depression and anxiety risk values.
• Figure 6: Typical non-biased sampling of coordinates within a neighbourhood in the final dataset.
• Figure 7: Examples of sampling that overrepresents certain parts of the neighourhoods to a different extent.