聽診器聲學特性分析與臨床呼吸音識別演算法

Abstract

本研究主旨在建立可量化聽診器收音品質的聲學測試系統，以及臨床呼吸異常音識別演算法。聽診器是醫學診斷上不可或缺的工具，電子聽診器的問世解決了聽診聲音無法收集、判斷只能仰賴資深醫療人員的問題。使用電子聽診器可讓錄製後的聲音訊號儲存在電腦，供進一步的分析與醫療決策之用，如何量化收音品質，提高電子聽診器的可靠度便成為重要的議題。 目前需要一套完整且科學化的聲學測試系統，來測試電子聽診器收音訊號的品質，本研究利用3M電子聽診器及連續式聽診貼片於無響室錄音，藉由播放單頻訊號並分析其頻率響應與諧波失真率，建立聲學特性量測系統；再透過臨床試驗，實際收集加護病房及呼吸照護病房病患之呼吸聲音。最後，利用機器學習方法建構一套呼吸音識別演算法，其中包含梅爾倒頻譜的特徵擷取、並比較三種類神經網路分類模型，用以自動判斷正常及異常呼吸音，增加臨床實際應用價值。 本研究中，聲學測試系統能有效地評估電子聽診器之收音品質、區分3M電子聽診器和連續聽診貼片之間的收音質量。結果顯示，3M電子聽診器因其構造及錄音軟體的既有設定性質，導致錄製後的聲音因為其頻率響應非直線穩定、分類模型較難提取特定的頻率特徵而降低辨識正確率。最後，我們的臨床呼吸音分類模型具有93％的總體準確度，可以在醫院中自動識別不同的呼吸聲並幫助醫生進行診斷。 未來將使用此聲學測試系統應用在聽診器開發與評估上，並持續優化臨床呼吸音識別演算法，納入更多臨床呼吸異常音種類，並建構一個可以連續監控、即時警示的遠距智慧呼吸照護系統。

Stethoscope is an indispensable tool for medical diagnosis. Electronic stethoscope solves the problem with traditional stethoscope, auscultation cannot be recorded and stored, and that diagnosis can only rely on medical professionals. By using electronic stethoscopes, sound signals are now transmittable to computers for further analysis. Therefore, quantification of the recording quality of electronic stethoscopes has become an important indicator. The aim of this study is to establish an acoustic testing system which quantifies the recording quality of electronic stethoscopes, and to develop a clinical respiratory sound classification algorithm. Currently, there are no complete and scientific standards for evaluating stethoscopes. In this study, monotonic sound is played using the frequency sweep method and recorded by 3M electronic stethoscopes and continuous auscultation patches in a soundless room. After calculating frequency response and harmonic distortion rate to quantify acoustical properties, an acoustic testing system is developed. Through clinical trials in ICU and RCW in hospitals, patients’ respiratory sound is collected. Machine learning algorithm is then used to build up a respiratory sound recognition and classification system, including MFCC feature extraction and selecting the best out of 3 types of neural networks models, to automatically recognize normal and abnormal respiratory sound. In this study, the acoustic testing system works effectively on evaluating an electronic stethoscope, enabling it to differentiate on the quality of recording between 3M electronic stethoscopes and continuous auscultation patches. The result shows that because of 3M stethoscope’s structure and recoding restrictions, sound recorded by it has unstable frequency response which leads to the classification model unable to extract the specific frequency features and lower the accuracy of recognition. This difference in recording quality of stethoscope correlates to the classification performance. Finally, our classification model has an overall accuracy of 93%, which means that it can be used in hospitals to automatically recognize different respiratory sound and help doctors for diagnosis. In the future, the acoustic testing system will be used for developing and evaluating electronic stethoscope. In addition, respiratory sound recognition algorithm will be trained and will include more clinical abnormal sound types; furthermore, this will build up a continuous monitoring and real-time alerting respiratory caring system.