A spatio-temporal attention network-based visual recognition method for post-stroke depression

As one of the major disabling and fatal diseases worldwide, the prevalence of post-stroke depression (PSD), a complication of stroke, continues to rise, exacerbating the mental health burden of patients. Studies have shown that facial information of patients has become a major source of data for depression identification due to its high psychological relevance and good practicality in application. In this study, we used ResNet as the backbone structure, combined with face recognition and alignment methods to efficiently preprocess the facial video data captured by the camera. Further, we utilized the spatial attention module and the temporal attention module to extract and integrate the facial features, through which we enhanced the correlation of the facial video information across pixels and frames, thereby extracting more effective features, and by fully leveraging the feature extraction capabilities of ResNet alongside the advantages of spatiotemporal attention mechanisms, which boosted model performance in PSD recognition tasks. The algorithm has been experimented on a public depression dataset and achieved 6.01 MAE and 7.25 RMSE. Compared experimentally with related algorithms, this method has shown significant performance improvement.