Identification of Individualised Aspiration Mechanisms in Post-Stroke Dysphagia: A VFSS-Based Interpretable Machine Learning Approach.

BACKGROUND

Post-stroke dysphagia (PSD) significantly increases the risk of aspiration-related pneumonia. While prior studies have linked abnormal temporal coordination of swallowing motions to aspiration, limited efforts have been made to explore its individualised mechanisms. This study aims to identify the biomechanical mechanisms underlying aspiration in PSD by developing an interpretable machine learning model based on videofluoroscopic swallowing study (VFSS) data.

METHODS

A retrospective analysis was conducted on 135 PSD patients with dysphagia, comprising 547 individual swallowing events. Ten VFSS-based temporal parameters were extracted, with aspiration and non-aspiration used as binary outcome variables. Five machine learning models and the Shapley Additive Explanations (SHAP) method were employed to evaluate variable importance and their non-linear contributions to aspiration risk.

RESULTS

The support vector machine (SVM) demonstrated the best performance (Accuracy: 92.7%). SHAP analysis identified laryngeal vestibule closure reaction time (LCRT), oral transit time (OTT), laryngeal closure to upper oesophageal sphincter opening interval (LC-UES), upper oesophageal sphincter opening duration (UOD) and stage transition duration (STD) as the most predictive parameters. Further interpretation demonstrated that aspiration results from complex temporal dysregulation rather than delays in isolated actions. Key parameters exhibited nonlinear and interactive effects, highlighting diverse compensatory or maladaptive patterns in impaired swallows.

CONCLUSIONS

By integrating VFSS-derived temporal parameters with interpretable machine learning, this study elucidates aspiration mechanisms in PSD at both the population and individual swallow levels, thereby providing a mechanism-driven basis for precision-guided clinical interventions.