Abstract WMP68: Machine Learning Approach In Prediction Of Nasoenteral Tube In Acute Ischemic Stroke

Abstract

Introduction: Dysphagia occurs in about 20 to 50% of acute stroke patients, and it may persist longer than six months after stroke. Nasoenteral tube feeding (NTF) must be judiciously prescribed to avoid dysphagia complications and, at the same time, prevent its unnecessary usage, which is not free of adverse events. Objective: Our aim was to state independent predictive factors associated with dysphagia and nasoenteral tube feeding. Besides, we aimed to develop a prediction model for nasoenteral tube feeding through a machine learning modeling approach. Methods: This is a prospective cohort study. All consecutive ischemic acute stroke patients were included. All patients had phonoaudiological evaluation for dysphagia screening. Data analyzed included age, sex, Glasgow Coma Scale, NIHSS, Aspects score, Seattle comorbid index, Rankin scale at discharge, TOAST classification for stroke subtypes, presence of major stroke risk factors and data from CT scan of patients. Results: We studied 1101 acute stroke patients. Twenty-eight percent of stroke patients received NTF. They were older (p<0.001), had a more severe stroke (p<0.001), and presented consciousness disturbance and dysarthria more frequently (p<0.001 for both measures). These findings independently predicted nasoenteral tube feeding in acute stroke. The decision tree model disclosed a sensitivity of 75% and specificity of 87%, with 84% accuracy for predicting nasoenteral tube feeding. On the other hand, the artificial neural network predicted 83% accurately, disclosing a sensitivity of 70% and specificity of 86%. Conclusions: Dysphagia occurred in one third of cases. Older age, stroke severity, dysarthria, reduced conscious level at onset independently predict the need for nasoenteral tube. The tree decision model is an accurate tool for predicting NTF in acute ischemic stroke patients.