Cross-modal knowledge distillation for enhanced depression detection

Depression is a severe mental disorder characterized by high prevalence, high recurrence, high disability, and high mortality rates. Consequently, timely detection and treatment are crucial. In recent years, speech-based methods for depression detection have been widely applied in clinical diagnostics. This is largely because they feature simple data collection and provide a positive user experience. However, these methods are limited by their susceptibility to deception and low accuracy rates. In contrast, brain information can provide neurobiological markers for depression, offering greater objectivity and higher accuracy in detection. To enhance the accuracy of speech-based detection, we propose a cross-modal knowledge distillation method that integrates speech signals with brain information, resulting in a more reliable and universally applicable approach for detecting depression. Specifically, a new multimodal model was first constructed as the teacher model, while the unimodal speech model served as the student model. Through knowledge distillation, the rich knowledge encapsulated in the teacher model was transferred to the student model, thereby enhancing its performance. Finally, experiments were conducted on the public MODMA dataset, achieving an accuracy of 83.19% for the distilled model, which represents a 3.47% improvement compared to traditional speech recognition methods. These results validate the effectiveness and feasibility of our proposed method, offering more effective support for the clinical diagnosis of depression.