Using Speech Features and Machine Learning Models to Predict Emotional and Behavioral Problems in Chinese Adolescents

Background: Current assessments of adolescent emotional and behavioral problems rely heavily on subjective reports, which are prone to biases. Aim: This study is the first to explore the potential of speech signals as objective markers for predicting emotional and behavioral problems (hyperactivity, emotional symptoms, conduct problems, and peer problems) in adolescents using machine learning techniques. Materials and Methods: We analyzed speech data from 8215 adolescents aged 12-18 years, extracting four categories of speech features: mel-frequency cepstral coefficients (MFCC), mel energy spectrum (MELS), prosodic features (PROS), and formant features (FORM). Machine learning models-logistic regression (LR), support vector machine (SVM), and gradient boosting decision trees (GBDT)-were employed to classify hyperactivity, emotional symptoms, conduct problems, and peer problems as defined by the Strengths and Difficulties Questionnaire (SDQ). Model performance was assessed using area under the curve (AUC), F1-score, and Shapley additive explanations (SHAP) values. Results: The GBDT model achieved the highest accuracy for predicting hyperactivity (AUC = 0.78) and emotional symptoms (AUC = 0.74 for males and 0.66 for females), while performance was weaker for conduct and peer problems. SHAP analysis revealed gender-specific feature importance patterns, with certain speech features being more critical for males than females. Conclusion: These findings demonstrate the feasibility of using speech features to objectively predict emotional and behavioral problems in adolescents and identify gender-specific markers. This study lays the foundation for developing speech-based assessment tools for early identification and intervention, offering an objective alternative to traditional subjective evaluation methods.