十二導極心電訊號合成之類神經網路法

Neural Network Methods for the Synthesis of 12-Lead ECG Signals

張雅昀(Ya-Yun Chang)

指導教授 ： 伍紹勳

國立交通大學/資訊學院/生醫工程研究所/碩士(2019年)

摘要

傳統十二導極心電圖儀需在身上黏貼十個導極，並且導線纏身，病人如果想要行動，較不方便，而醫院內也有可攜式動態心電圖儀侯特式心電圖儀系統，黏貼三至十二個導極在病人身上，但是需要背著一臺小機器在身上，對於長時間監控的患者會感受到機器重量的負擔。因此，為了增加心電圖監測的方便性，希望可以不受多導極黏貼的限制，將少導極的心電訊號合成完整十二導極，提供醫生心臟多面向的向量資訊，提升醫療品質。
本研究使用長短期記憶模型，藉由心電訊號時間與導極空間上的相關性找出三導極與十二導極合成之間的轉換矩陣；也使用雙向長短期記憶模型，不僅考慮前向的時間相關性，也考慮後向的時間相關性；以及使用變分自動編碼器架構，藉由將數據壓縮編碼到潛在空間，再解碼生成數據的訓練模型，找出一個更適用多人的三導極與十二導極間之通用轉換矩陣，讓心電訊號合成演算法可以更廣泛應用。最後，實做相關研究進行訊號合成結果相關性比較。
　　利用Biosemi公司的Active Two系統，在新光吳火獅醫院進行人體試驗，收集受試者身體表面電位共八十三位受試者的三十二個電極量測的心電訊號，並取得線上開放式心電訊號資料庫PTB共五十二筆，與美國食品與藥物管理局認可的CSE資料庫共一百筆。將每筆資料三導極與十二導極分割成五秒一段，放入訓練資料集以及測試資料集。
　　實驗結果長短期記憶模型三百二十八筆測試資料合成的十二導極平均相關係數為0.87，雙向長短期記憶模型的平均相關係數為0.87，變分自動編碼器模型的平均相關係數為0.9906。合成結果的心電圖波形特徵大多符合原量測心電訊號的特徵，且在擁有心臟疾病的資料合成結果也有達成同樣的效果。

關鍵字

健康照護 ； 心電圖 ； 心電圖合成 ； 類神經網路

並列摘要

The traditional twelve-lead electrocardiograph needs to attach ten electrodes to the body, and the wires are entangled. If the patient wants to move more conveniently, the hospital also has an ambulatory electrocardiograph called Holter System. Attaching three to twelve electrodes to the patient, but patients need to carry a small machine. Therefore, in order to improve the convenience of electrocardiograph monitoring, it is hoped that combine the advantages of traditional twelve-lead ECG and ambulatory ECG.
　　This study uses Long Short-Term Memory to find the transformation matrix of three differential leads and synthesized twelve-lead electrocardiograph signals by the correlation between the feature of time and space. The Bidirectional Long Short-Term memory model considers not only the temporal correlation of the forward direction but also the temporal correlation of the backwards. Variational Autoencoder by encoding the data in latent space, then decoding the training data to find a conversion matrix between the three-lead ECG and twelve-lead ECG is more suitable for many people.
　　Using Biosemi's Active Two system, we conducted clinical studies and experiments at Shin Kong Wuhuoshi Hospital to collect 83 subjects' body surface potential mapping data and obtained 52 data of PTB database, and 100 data from CSE databases approved by the US FDA. Dividing each three differential leads and twelve-lead data into five-second segments, and putting them into the training data set and the test data set.
　　The average correlation coefficient of the twelve-lead synthesized by LSTM model is 0.87, the average correlation coefficient of the BLSTM model is 0.87, and the average correlation coefficient of the VAE model is 0.9906. Most of the ECG waveform features of the synthesis result are consistent with the original ECG signal, and the same effect can be achieved in the synthesis of some heart diseases.

並列關鍵字

Telecardiology ； ECG ； Synthesis of 12-lead ECG ； Neural Network

參考文獻

[1] Barold, S., “Willem Einthoven and the Birth of Clinical Electrocardiography a Hundred Years Ago.,” CEPR, Vol. 7, No. 1, pp.99-104, Jan. 2003.

[2] Jin, Benjamin E et al., “A simple device to illustrate the Einthoven triangle,” Advances in physiology education, vol. 36, pp. 319-324, Dec. 2012.

[3] CDC. (2008, Feb. 10). Disparities in Adult Awareness of Heart Attack Warning Signs and Symptoms[Online]. Available: https://www.cdc.gov/mmwr/preview/mmwrhtml/mm5707a3.htm

[4] Roberts, William C. et al., “Norman Jefferis Holter and ambulatory ECG monitoring.,” Am J Cardiol, vol. 52, pp.903-906, Oct. 1983.

[5] Costin, Hariton et al.,“Short-term Heart Rate Variability using wrist-worn pulse wave monitor compared to a Holter ECG.,” in EHB, Sinaia, 2017, pp. 635-638.