Machine learning-based analysis of the relationship between brain lesion sites and swallowing and cognitive functions in stroke patients

Abstract

Abstract Stroke-related dysphagia is influenced by brain damage location and cognitive impairment, but its mechanisms remain unclear. In this study, we aimed to clarify the mechanisms by which brain damage causes dysphagia and cognitive function in 246 patients with stroke using a probabilistic neural network model. Dysphagia was classified as mild (oral intake with liquid and diet modifications) or severe (unable to take food orally, requiring tube feeding). Atlas-based segmentation applied to brain MRI data delineated 121 anatomically-defined regions, including 116 Automated Anatomical Labeling regions (AAL116), brainstem level 1, and white matter level 4 of the Automated Talairach Atlas Labels (ATAL), and the total score of five cognitive items on the Functional Independence Measure was used to evaluate cognitive function. Classifying dysphagia by severity and evaluating cognitive function resulted in improvements in prediction accuracy and reduced the number of predictor variables. In addition, after adding evaluations of cognitive function in both the severe and mild dysphagia groups, evaluations of the brainstem, which had remained a final predictor variable in the analysis of brain regions only, no longer remained. The results highlight the importance of integrating neurological imaging and cognitive assessment in the diagnosis and rehabilitation of dysphagia after stroke.