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

Area of Science:

• Adolescent psychology
• Machine learning applications
• Speech signal processing

Background:

• Current adolescent emotional and behavioral problem assessments rely on subjective, potentially biased reports.
• Objective markers are needed to improve diagnostic accuracy and reduce bias.

Purpose of the Study:

• To investigate speech signals as objective predictors of adolescent emotional and behavioral problems.
• To explore machine learning techniques for classifying hyperactivity, emotional symptoms, conduct problems, and peer problems using speech data.
• To identify gender-specific speech features associated with these problems.

Main Methods:

• Analysis of speech data from 8215 adolescents (aged 12-18).
• Extraction of speech features including mel-frequency cepstral coefficients (MFCC), mel energy spectrum (MELS), prosodic features (PROS), and formant features (FORM).
• Application of machine learning models (logistic regression, support vector machine, gradient boosting decision trees) for classification, with performance evaluated by AUC, F1-score, and SHAP values.

Main Results:

• Gradient Boosting Decision Trees (GBDT) model showed highest accuracy in predicting hyperactivity (AUC=0.78) and emotional symptoms (AUC=0.74 for males, 0.66 for females).
• Model performance was less robust for conduct and peer problems.
• Shapley Additive Explanations (SHAP) analysis highlighted gender-specific importance of certain speech features.

Conclusions:

• Speech features can objectively predict emotional and behavioral problems in adolescents.
• Identified gender-specific speech markers offer potential for targeted interventions.
• This research supports the development of speech-based tools for early identification, complementing traditional subjective methods.