Transfer Learning for Dysphagia Detection

This study evaluates the use of transfer learning techniques to differentiate normal and pathological swallowing sounds. Audio data were obtained by recording swallow sounds from healthy participants and patients with dysphagia using a commercially available wearable stethoscope. In our previous study which used a traditional approach, statistically significant differences were found in acoustic features such as spectral crest and zero-crossing rate, between normal and dysphagic swallows. Random forest modelling resulted in fair sensitivity (mean ± SD: 74% ± 8%) and specificity (89% ± 6%) for dysphagic swallows, with an overall accuracy of 83% ± 3% and an F1 score of 78% ± 5%. To explore whether a deep learning approach would improve classification performance, transfer learning methods (Vggish, Yamnet, and OpenL3) were employed. These methods significantly outperformed the random forest model used in our previous study. Vggish achieved the highest metrics with 91.52% precision, 87.65% sensitivity, 93.70% specificity, 91.13% accuracy, and an F1 score of 89.38%. Yamnet (Precision: 90.80%, Sensitivity: 88.24% Specificity: 93.26%, Accuracy: 91.13%, F1 score: 89.25%) and OpenL3 (Precision: 87.68%, Sensitivity: 87.37% Specificity: 90.66%, Accuracy: 89.65%, F1 score: 87.35%) also showed superior performance with precision and accuracy metrics. Overall, these results demonstrate that cross-domain transfer learning approaches can be valuable tools in enhancing the accuracy and reliability of dysphagia detection, and the need for further research to optimize these techniques for clinical use.