Development and Validation of an Interpretable Machine Learning Model for Predicting Tic Disorders and Severity in Children Based on Electroencephalogram Data

Accurate diagnosis of Tic disorders (TD) and its severity based on electroencephalogram (EEG) data were of great clinical importance. This study analyzed EEG data from 90 children with TD and 88 healthy controls (HC). A two-stage progressive diagnosis framework based on EEG data and machine learning methods was developed. To achieve individualized prediction and reduce the feature dimension, we proposed a novel individual-based feature-weighted integration method in machine learning, as well as a new SHAP-driven feature selection and weighting (SFSW) strategy to improve the prediction accuracy. Based on 13 weighted features, Logistic Regression model achieved an average accuracy of 94.2% (95% CI, 90.6%-97.9%) in diagnosing TD, with a sensitivity of 92.4% (95% CI, 85.3%-99.5%) and a specificity of 96.1% (95% CI, 92.9%-99.2%). The Decision Tree model attained an average accuracy of 81.5% (95% CI, 68.6%-94.5%) in predicting severity, with a sensitivity of 81.5% (95% CI, 68.6%-94.5%) and a specificity of 89.9% (95% CI, 82.1%-97.6%). In the hold-out set validation, the method demonstrated accuracy rates of 95.7% in diagnosing TD and 83.3% in predicting severity. Interpretability analysis revealed that the top three main features affecting TD diagnosis were the mean frequency (MNF) of P3 channel $\beta $ band, age and MNF of C3 channel $\gamma $ band. This work offered a more efficient approach to individualized diagnosis of TD and had substantial practical value for clinical auxiliary diagnosis and intervention.