Personalized prediction and intervention for adolescent mental health: multimodal temporal modeling using transformer

TL;DR

This paper introduces MPHI Trans, a new model that uses multimodal data and temporal modeling to predict and intervene in adolescent mental health more accurately.

Contribution

The novel MPHI Trans model improves mental health prediction by integrating multimodal data and temporal modeling.

Findings

MPHI Trans outperformed advanced models like BERT and T5 on DAIC-WOZ and WESAD datasets.

Temporal modeling and multimodal fusion modules are critical for capturing emotional fluctuations.

The model achieved high accuracy, recall, and AUC-ROC scores on benchmark datasets.

Abstract

Adolescent mental health problems are becoming increasingly serious, making early prediction and personalized intervention important research topics. Existing methods face limitations in handling complex emotional fluctuations and multimodal data fusion.

To address these challenges, we propose a novel model, MPHI Trans, which integrates multimodal data and temporal modeling techniques to accurately capture dynamic changes in adolescent mental health status.

Experimental results on the DAIC-WOZ and WESAD datasets demonstrate that MPHI Trans significantly outperforms advanced models such as BERT, T5, and XLNet. On DAIC-WOZ, MPHI Trans achieved an accuracy of 89%, recall of 84%, precision of 85%, F1 score of 84%, and AUC-ROC of 92%. On WESAD, the model attained an accuracy of 88%, recall of 81%, precision of 82%, F1 score of 81%, and AUC-ROC of 91%.

Ablation studies confirm the critical contributions of the temporal modeling and multimodal fusion modules, as their removal substantially degrades model performance, underscoring their indispensable roles in capturing emotional fluctuations and information fusion.

Figures