Voice-based machine learning for rapid screening of bipolar disorder and major depressive disorder in children and adolescents: a robust and low-complexity diagnostic model

Background: Major depressive disorder (MDD) and bipolar disorder (BD) are psychiatric disorders that seriously impact physical and mental health. They are increasingly prevalent among children and adolescents, and the absence of objective physiological indicators makes diagnosis difficult. However, existing studies have primarily focused on adults, and few practical diagnostic tools have been developed and clinically deployed. Therefore, we investigated the voice features of children and adolescents and proposed a low-complexity automatic detection method for early recognition and self-screening. Methods: A reading paradigm with 7 segments of text is applied for voice data collection. After dividing the recording, a well-developed feature set is extracted, and the double feature selection method is proposed to select the most effective features. Finally, traditional classification models are applied to reduce complexity. Results: The energy, spectral slope, amplitude spectrum, and RASTA-style filtered auditory spectrum of voice are effective features. Results show that 92.4% and 95.6% for voice and subject accuracy are achieved in the ternary classification of 50 BD, 50 MDD, and 50 healthy controls (HC). Besides the satisfactory accuracies, the robustness to recording devices and environments is validated. Conclusions: Voice features are potential biomarkers for diagnosing psychiatric disorders in children and adolescents. Based on optimized feature selection algorithms, traditional classifiers can achieve accurate and robust classification of BD, MDD, and HC with a small number of interpretable features, providing a feasible tool for auxiliary diagnosis.