基於壓縮分析之有效心律不整偵測機制於遠程心電圖監護系統

Abstract

為了落實心電圖的遠程監護系統，壓縮感知 (compressive sensing) 是近年來被提出的新技術，用於降低生理感測器及資料傳輸的功耗。在家中，病患的生理訊號可由壓縮感知的感測器量取，並傳送到醫院或是手機來達到遠程監護。當壓縮訊號傳送到後端時，必須先由還原演算法將原始波形重建。心電圖遠程監護系統的目的在於，透過自動偵測異常狀況來減輕醫生的負擔。醫院的雲端伺服器可先挑出高危險群的病患再由醫生診斷；在家中，手機可對異常狀況即時發出警告並提供緊急措施。然而，醫院的伺服器及手機分別存在著資料量的可擴展性及運算資源受限的問題。因此，我們的目標為開發適用於心電圖壓縮遠端監護系統的低複雜度壓縮分析(compressed analysis)架構。 由於壓縮感知還原過程的複雜度比起其他運算高，改變還原過程可有效達到低複雜的系統。在醫院，為了提供醫生完整的心電圖波形，還原是必須的。因此我們希望利用壓縮感知的特徵，在還原的過程中達到自動偵測。透過將還原的過程拆成兩個步驟，我們可以在第一步驟先用低複雜度的運算偵測，而只有高危險的訊號需要在第二階段做完整的還原。如此一來，系統的複雜度可以降低且同時達到相同的成效。另一方面，在資源受限的手機上，我們提出一壓縮分析的架構，利用主成份分析(principle component analysis)協助找到的字典，直接將壓縮訊號轉換到特徵空間，由此擷取出偵測所需的資訊而不經過還原的過程，使得整個系統的複雜度得以降低。總結來說，提出的壓縮分析架構可以同時達到降低複雜度及維持系統成效。

To realize Electrocardiography (ECG) telemonitoring systems, compressive sensing (CS) is a new technique to reduce power of biosensors and data transmission. Patients can stay at home, and their biosignals can be measured by CS sensors. These compressively sensed signals are transmitted to hospital or cell phones for remote monitoring. The CS reconstruction is required to obtain the original waveform on the receivers. The purpose of the ECG telemonitoring system is to automatically detect abnormal situation in order to release the load of doctors. In the hospital, the cloud servers can first select high-risks patients for doctors. At home, the users’ cell phone can real-time alert on abnormal situations and provide some emergency procedures. However, there exist scalability and resource issues on the servers in hospital and the users’ cell phones respectively. As the consequence, our goal is to develop a low-complexity compressed analysis framework for the CS-based ECG telemonitoring system. As the overhead of CS reconstruction is higher than other computation, the low-complexity system can be achieved by modifying this process. In hospital, the CS reconstruction is required to provide the original waveform to doctors. Therefore, we exploit the features of CS to detect during the reconstruction process. By decomposing the reconstruction process into two stages, signals are detected with low overhead in the first stage; and only high-risks signals are fully reconstructed in the second stage. In this way, the system can achieve low complexity with the same performance. On the other hand, for the resource constrained nodes like cell phones, we propose a compressed analysis framework using principle component analysis-assisted dictionary to directly extract information from compressed signals without reconstruction. With this framework, the system complexity can be reduced by removing the reconstruction process. In summary, the proposed compressed analysis frameworks can achieve lower complexity and remain the same performance in both scenarios.