使用捲積神經網路分類吞嚥聲學訊號的聲譜圖以區分有無吞嚥障礙者的方法

// Method for Classifying Spectrograms of Acoustic Signals from Swallowing of Individuals with and without Dysphagia Using a Convolutional Neural Network

本研究以頸部聽診（cervical auscultation）與數位聽診器錄得178次吞嚥聲並擴增至1,888個事件，轉為STFT時頻圖後以卷積神經網路（CNN）交叉驗證訓練｜模型表現為準確率77.7%、敏感度76.6%、精確度64.1%、AUC=0.823，顯示以神經網路分析吞嚥聲可作為傳統吞嚥障礙評估的輔助篩檢工具，但仍需進一步驗證穩健性與臨床可行性。

// This study developed and tested a machine-learning method to classify dysphagia from cervical auscultation recordings. Researchers recorded 178 swallowing sounds with a digital stethoscope, augmented them to 1,888 events, converted signals to STFT spectrograms, and trained a convolutional neural network with cross-validation. The model reached 77.7% accuracy, 76.6% sensitivity, 64.1% precision, and an AUC of 0.823. Results suggest neural-network analysis of swallowing sounds can complement traditional dysphagia assessment, though robustness and clinical applicability require further work.

Published 2026年9月11日