Spatiotemporal Electrical Impedance Tomography for Speech Respiratory Assessment in Cleft Palate: an Interpretable Machine Learning Study

Objective Cleft palate (CP) is a common congenital deformity often associated with velopharyngeal insufficiency (VPI), which disrupts the physiological coupling between respiration and speech. Conventional clinical assessments, such as nasometry and spirometry, provide limited static data and fail to visualize the dynamic spatiotemporal distribution of lung ventilation during phonation. This study introduces spatiotemporal electrical impedance tomography (ST-EIT) to evaluate speech-respiratory functional features in CP patients compared to normal controls (NC). The aim is to characterize multi-domain respiratory patterns and to validate an interpretable machine learning framework for providing objective, quantitative evidence for clinical assessment. Methods Seventy-five participants were enrolled in this study, comprising 37 patients with surgically repaired CP and 38 healthy volunteers matched for age, gender, and body mass index (BMI). All subjects performed standardized sustained phonation tasks while undergoing synchronous monitoring with a 16-electrode EIT system and a pneumotachograph. A comprehensive feature engineering pipeline was developed to extract physiological parameters across 3 complementary domains. (1) Temporal domain: including inspiratory/expiratory phase duration (tPhase), time constants (Tau), and inspiratory-to-expiratory time ratios (TI/TE); (2) airflow domain: comprising mean flow, peak flow, and instantaneous flow at 25%, 50%, and 75% of tidal volume; and (3) spatial domain: quantifying global and regional tidal impedance variation (TIV), global inhomogeneity (GI), and center of ventilation (CoV). Extreme Gradient Boosting (XGBoost) classifiers were trained using 5 distinct data sources (Spirometry, Nasometry, Inspiratory-EIT, Expiratory-EIT, and fused ST-EIT). Model performance was rigorously evaluated via stratified 5-fold cross-validation, and Shapley additive explanations (SHAP) were employed to quantify global and local feature contributions. Results The CP group exhibited a distinct respiratory phenotype compared to controls. In the temporal domain, CP patients showed significantly shorter inspiratory (1.60 s vs. 1.85 s, Pvs. 3.95 s, PPAUC: 0.915±0.012, Accuracy: 0.843±0.019, F1-score: 0.872±0.017), substantially outperforming models based on spirometry (AUC: 0.721) or nasometry (AUC: 0.625) alone. Interpretability analysis revealed that spatial domain features were the most critical, contributing 53.4% to the model’s decision-making, followed by temporal (25.0%) and airflow (21.6%) features. Conclusion ST-EIT successfully captures the temporal, airflow, and spatial deviations in CP speech respiration that are undetectable by conventional methods—specifically, rapid phase transitions, hyperdynamic expiratory airflow, and regional ventilation heterogeneity. This study validates ST-EIT as a robust, non-invasive, and radiation-free tool for characterizing speech-respiratory dysfunction, offering high clinical value for bedside screening, rehabilitation planning, and longitudinal monitoring of patients with cleft palate.