A Deep Learning-Based Decision Support System for Early Depression Screening from Multimodal Clinical Interviews

Abstract

The complexity of the behavioral, emotional, and linguistic features of depressive disorders is still a challenge for depression screening by clinical interviews. Self-report questionnaires, like the PHQ-9, have a number of limitations, including self-report bias and inadequate behavioral representation. In this paper, we thus propose the development of a Deep Learning-Based Decision Support System (DL-DSS) for early depression screening from multimodal clinical interview data. The proposed framework combines the transformer-based text encoding, CNN-based acoustic and visual feature extraction, and LSTM-based temporal modeling approaches to perform multi-modal analyses of transcript, speech, and facial behavior. The system was tested with the DAIC-WOZ benchmark data set. Experimental results show that the proposed multimodal DL-DSS model achieved 87.9% classification accuracy, 89.0% F1-scores, and 92.0% AUC-ROC, outperforming the conventional PHQ-9-based and single-modality deep learning approaches, and also the proposed model has a real-time inference latency of less than 120 ms per interview segment. These results lend support to the effectiveness of AI-based multimodal deep learning for intelligent and scalable depression screening within current clinical decision-support systems.