適用於穿戴式醫療的十二導程心電圖合成系統

12-lead ECG Synthesis for Wearable Health Device

黃于倢(Yu-Chieh Huang)

指導教授 ： 張文輝 ； 王逸如

國立交通大學/電機學院/電信工程系所/碩士(2019年)

摘要

因應高齡化社會的時代趨勢，遠距健康照護已成為先進國家重點發展的新興服務產業。本論文旨在探討穿戴式心臟照護系統的兩項關鍵技術，一對十二導程的心電圖合成及其在自動偵測心肌梗塞的應用。在第一項研究中，我們提出以穿戴式裝置量測所得的單一肢導程I為基礎之兩階段合成架構。第一階段利用輸入時延神經網路以及長短期記憶網路進行I對II、V2的一對三導程心電圖合成，第二階段再利用前饋式神經網路進行三對十二導程的心電圖合成。在第二項研究中，我們結合I導程與合成的胸前導程進行四種心肌梗塞病症的自動偵測。其關鍵在於針對每個心跳周期的前後兩部分進行多項式擬合的特徵擷取，再透過前饋式神經網路分類器判定受測者為正常或罹患前壁、前側壁、下壁、下側壁的心肌梗塞。實驗結果顯示，以長短期記憶網路為基礎的一對三導程心電圖合成效果最好，且忠實呈現了醫生臨床診斷所需的病理特徵。針對PTB Diagnostic資料庫50位患者進行的心肌梗塞測試中，使用I導程的正準率僅有74.56%，加入合成的胸前導程大幅提升其正確率最高達到98.59%。

關鍵字

穿戴式心臟照護 ； 心電圖合成 ； 心肌梗塞 ； 前饋式神經網路 ； 輸入時延神經網路 ； 長短期記憶網路

並列摘要

Wireless telecardiology is an active research with the goal of ubiquitous heart-care services. This work investigates two important aspects of a wearable heart-care system: ECG synthesis and Myocardial Infarction (MI) detection. We target the single-lead ECG configuration that is routinely used in wearable device. The first part of this work focuses on the patient-specific 12-lead ECG synthesis from the limb lead I ECG using a two-step procedure. In the first step, ECG synthesis of lead II and V2 from lead I is carried out by using input delay neural network (IDNN) and long short-term memory (LSTM) network. In the second step, lead I and synthesized lead II and V2 are used to reconstruct the 12-lead ECG using the Feedforward neural network (FNN). Experimental results show that the proposed method can synthesize the 12-lead ECG from single-lead ECG with an insignificant loss of diagnostic information. The second part of this wok presents the combined use of limb lead I and one synthesized precordial lead for automatic detection and localization of four types of MI, including Anterior, Anterior-Lateral, Inferior and Inferior-Lateral. The first step toward realization is to extract MI-related features by dividing a beat cycle into two parts and fitting each part with a polynomial function. After feature extraction, a FNN classifier is used to derive the most likely MI of an unknown subject. The results indicate that with the aid of ECG synthesis the system design can be developed to better exploit the intrinsic natures of MI morphologies. Experiments on the PTB Diagnostic database demonstrates that the proposed two-lead system predicts 50 subjects with an accuracy reaching 98.59%, compared to 74.56% of the single-lead system.

並列關鍵字

wearable heart-care ； ECG synthesis ； Myocardial Infarction (MI) ； feedforward neural network ； input delay neural network ； long short-term memory

參考文獻

[1] World Health Organization, https://www.who.int/en/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds).

[2] Stefan P. Nelwan, Jan A. Kors, Simon H. Meij, Minimal lead sets for reconstruction of 12-lead electrocardiograms. Journal of Electrocardiology, vol. 33, pages 163-166, 2000.

[3] H. Atoui, J. Fayn and P. Rubel, "A Novel Neural-Network Model for Deriving Standard 12-Lead ECGs From Serial Three-Lead ECGs: Application to Self-Care," in IEEE Transactions on Information Technology in Biomedicine, vol. 14, no. 3, pp. 883-890, May 2010.

[4] M. H. Ostertag and G. R. Tsouri, "Reconstructing ECG precordial leads from a reduced lead set using independent component analysis," 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 4414-4417, 2001.

[5] F. M. Al-Naima, A. H. Ali and S. S. Mahdi, "Data acquisition for myocardial infarction classification based on wavelets and Neural Networks," 2008 5th International Multi-Conference on Systems, Signals and Devices, Amman, pp. 1-6, 2008.