Speech-based early depressive symptom screening using MAX-Net: a multi-scale autoencoder-augmented xLSTM framework.

Background: Early screening for depressive symptoms lacks objective and stable biomarkers, while traditional methods based on questionnaires and subjective assessments suffer from issues such as low efficiency. Although the use of artificial intelligence to analyze speech signals has become a growing trend, existing methods are limited by challenges such as small medical datasets, complex features, and long-term temporal dependencies, resulting in insufficient recognition accuracy and poor cross-dataset generalization performance. Methods: To address the above issues, this paper proposes MAX-Net, a multi-scale feature-enhanced long-term sequential autoencoder network based on an improved extended long short-term memory (xLSTM) for depressive symptom screening. MAX-Net combines the improved xLSTM with autoencoder feature enhancement and introduces a Multi-Scale Convolutional layer (MSC) module. To mitigate the issue of data sparsity, we propose an autoencoder-based feature space enhancement strategy. Furthermore, we investigate the impact of varying numbers of augmented samples on model performance. The model is trained and evaluated on the in-house dataset HMU-DDAC-24 and retrained on the public dataset MODMA using the same architecture and hyperparameters to assess its stability and cross-dataset applicability. Results: Experiments on HMU-DDAC-24 show that MAX-Net achieves an Area Under the ROC Curve (AUC) of 0.945 and a Mean Absolute Error (MAE) of 1.61 in the depression symptom classification task, significantly outperforming existing baseline models. It maintains competitive performance on the MODMA dataset (AUC = 0.804), suggesting architectural stability and consistent effectiveness across different data sources when retrained. Further experiments revealed that the optimal data augmentation strategy varies across different datasets: on HMU-DDAC-24, performance peaks when 15,000 samples are generated, whereas on MODMA, performance stabilizes when 18,000-22,000 samples are generated. Conclusion: MAX-Net provides an effective framework for speech-based early depression screening and demonstrates consistent performance across datasets under a unified configuration. These findings indicate that the proposed architecture is robust to variations in data distribution. Furthermore, the autoencoder-based augmentation strategy has proven beneficial in low-data-volume scenarios. Overall, this study supports the application potential of MAX-Net as an auxiliary tool for early depression screening.