Psychological crisis detection based on behavioral data: a new approach to non-contact measurement

Timely identification of psychological crisis reduces the risk of self-harm or suicide and maintains mental health. Previous methods of identifying psychological crisis have relied on contact methods such as annual psychological surveys and voluntary reports from peers and teachers to identify students in psychological crisis. However, these methods have the disadvantages of time lag and limited scope. This study aims to propose a contactless measurement method that can detect psychological crises in a timely manner. Based on multi-dimensional data, this study constructed a feature set of psychological crises through digital phenotyping, established prediction models of multiple machine learning algorithms, and verified their effectiveness.

We collected behavioral data from 4,508 college students based on campus smart cards. Based on this, we constructed a psychological crisis prediction feature set covering five major categories and 39 sub-features, including interpersonal relationship, academic performance, physical activity and health, economic level, and demographic characteristics. The correlation between each sub-feature and psychological crisis was analyzed. Then, the classification algorithms of Random Forest, Deep Neural Network, and Graph Convolutional Network were used to classify and predict the psychological crisis and to compare the performance of the algorithms under the balanced and unbalanced strategies. Finally, Top-N is introduced to provide actionable prioritization for actual screening.

The results showed that frequency of dining with classmates and roommates were negatively correlated with psychological crisis, while frequency of failed subjects and length of time staying up late were positively correlated with psychological crisis. In algorithm comparisons, we found that models without balancing strategies achieved higher accuracy, while those with balancing strategies demonstrated higher recall rates and balanced accuracy. Additionally, Random Forest performed best on the Top-N evaluation.

The findings suggest that the non-contact method is effective in predicting student psychological crisis. The use of Random Forest algorithm combined with Top-N assessment can effectively identify high-risk individuals and optimize intervention priorities. Furthermore, this study explored the applicability of non-contact methods for predicting students' psychological crisis in real-world scenarios.