A Machine Learning-Based Screening Test for Sarcopenic Dysphagia Using Image Recognition

Background: Sarcopenic dysphagia, a swallowing disorder caused by sarcopenia, is prevalent in older patients and can cause malnutrition and aspiration pneumonia. This study aimed to develop a simple screening test using image recognition with a low risk of droplet transmission for sarcopenic dysphagia. Methods: Older patients admitted to a post-acute care hospital were enrolled in this cross-sectional study. As a main variable for the development of a screening test, we photographed the anterior neck to analyze the image features of sarcopenic dysphagia. The studied image features included the pixel values and the number of feature points. We constructed screening models using the image features, age, sex, and body mass index. The prediction performance of each model was investigated. Results: A total of 308 patients participated, including 175 (56.82%) patients without dysphagia and 133 (43.18%) with sarcopenic dysphagia. The area under the receiver operating characteristic curve (ROC-AUC), sensitivity, specificity, positive predictive value, negative predictive value, and area under the precision-recall curve (PR-AUC) values of the best model were 0.877, 87.50%, 76.67%, 66.67%, 92.00%, and 0.838, respectively. The model with image features alone showed an ROC-AUC of 0.814 and PR-AUC of 0.726. Conclusions: The screening test for sarcopenic dysphagia using image recognition of neck appearance had high prediction performance.

Automated summary

This study developed a simple screening test using image recognition of neck appearance to detect sarcopenic dysphagia, which showed high prediction performance with an ROC-AUC of 0.877. The model with image features alone also demonstrated good prediction performance with an ROC-AUC of 0.814.