Artificial intelligence for dysphagia screening: A machine learning approach

Early and reliable screening for oropharyngeal dysphagia symptoms in at-risk populations is a crucial first stage in effective management. This study aims to propose a machine learning model that can predict whether individuals are at risk of having confirmed dysphagia based on diagnostic accuracy. This was an exploratory cross-sectional diagnostic accuracy study. Adults and older adults were screened and assessed, and a videofluoroscopy swallowing study was performed to confirm dysphagia in participants with complaints about swallowing and signs or symptoms identified in the clinical assessment by a speech therapist. Eight algorithms were tested based on a theoretical model to evaluate their diagnostic accuracy: Naive Bayes, K-Nearest Neighbors, Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, Multilayer Perceptron, and Convolutional Neural Networks. The Pandas library in Python was used for data visualization and manipulation. In total, 465 participants were included, of whom 153 had dysphagia. A screening with 15 items related to personal, health, and oral health characteristics is proposed. The final four best-performing models were Naive Bayes (accuracy 0.76, sensitivity 0.84), Logistic Regression (accuracy 0.79, sensitivity 0.82), Support Vector Machine (accuracy 0.78, sensitivity 0.79), and Convolutional Neural Network (accuracy 0.78, sensitivity 0.74). The use of artificial intelligence based on different machine-learning approaches is promising for dysphagia screening, demonstrating adequate diagnostic accuracy. Future studies should be conducted to externally validate the proposed screening models.