Predicting Long-Term Prognosis of Poststroke Dysphagia with Machine Learning

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Poststroke dysphagia is a common complication impacting quality of life.
• While many recover, some experience persistent swallowing difficulties beyond six months.
• Predicting long-term outcomes is crucial for patient management.

Purpose of the Study:

• To investigate the efficacy of machine learning in predicting long-term poststroke dysphagia prognosis.
• To utilize early videofluoroscopic swallowing study (VFSS) data for predictive modeling.

Main Methods:

• Retrospective analysis of VFSS data (within 1 month of stroke) and 6-month swallowing status.
• Selection and scoring of 14 key VFSS parameters.
• Application of five machine learning algorithms (Random Forest, CatBoost, KNN, LGBM, XGBoost) combined via ensemble methods.

Main Results:

• A dataset of 448 patients was utilized (70% training, 30% testing).
• The final ensemble model achieved high performance metrics: 0.98 accuracy, 0.94 precision, 0.84 recall, 0.88 F1-score, and 0.99 AUC.
• Demonstrated significant predictive power for long-term dysphagia prognosis.

Conclusions:

• Machine learning models effectively predict long-term poststroke dysphagia prognosis using early VFSS data.
• These models offer valuable predictive information for clinical decision-making.
• Early identification of persistent dysphagia can guide timely interventions and support.