基於深度學習的短時語音錄音穿透-誤吸事件聲學篩檢

// Deep Learning-Based Acoustic Screening for Penetration-Aspiration Events Using Short Voice Recordings

本研究評估一款手機應用的深度學習人工智能工具，透過吞嚥前後的短語音錄音偵測吞嚥後氣道受損｜採用自動編碼器異常檢測模型訓練正常吞嚥聲，並以攝影透視吞嚥檢查驗證｜結果顯示模型靈敏度約91%、準確率85-90%，內部性能優異（AUC最高0.98），提供簡便可行的篩檢方法，有助及早發現需進一步吞嚥功能評估者。

// This study evaluated a smartphone-based deep learning AI tool to detect post-swallow airway compromise using brief voice recordings before and after swallowing. An autoencoder anomaly detection model was trained on normal swallowing sounds and validated against videofluoroscopic swallowing studies. The model showed high sensitivity (around 91%) and good accuracy (85-90%) with strong internal performance (AUC up to 0.98). This method offers a practical, accessible way to screen individuals needing further swallowing assessment.

Published 2026年6月2日