適用於生理訊號監護之壓縮感知即時診斷及還原引擎

Abstract

隨著高齡化社會及慢性病增長的趨勢下，健康議題漸漸地受到重視。面對接踵而來的醫療照護需求，可攜式的生理訊號的穿戴式產品被視為遠端醫療照護中的重要環節，全球許多研發團隊亦投入這方面的研究。當使用者將這些穿戴式裝置穿在身上時，感測器將會蒐集人體的生理訊號並且將相關數據傳送到後端的行動裝置或是雲端做進一步的分析。以心電訊號感測裝置為例，由於心電訊號已被證實能用於診斷許多疾病，並且透過感測器長時間蒐集心電訊號，因此心電訊號的監測被廣泛的應用在行動看護、居家照護、運動生理監測上。此外，隨著後端的疾病診斷分析系統越來越成熟，這些生理訊號可以被有效的分析並且做出精準的疾病診斷與偵測，進而改善人類的生活品質。 大多數的穿戴式裝置都有電池續航力、存儲容量有限以及受限的運算能力的問題。因此，如何發展出極度輕量化的穿戴式裝置，使其能夠減少訊號採集的複雜度、提高數據傳輸的效率將會是智能穿戴式裝置重要的課題。而壓縮感知 (compressed sensing，CS) 可以為智能穿戴所遇到的難題提供優良的解決方案。第一，壓縮感知透過一個隨機投影 (random projection) 將訊號採擷 (sampling) 和壓縮 (compression) 合而為一，如此能夠大幅降低感測器前端的複雜度； 第二，壓縮感知能夠以低於傳統奈奎斯頻率來取樣訊號，因此整體所需要傳輸和處理的的資料量可以大幅減少，以增加傳輸效率。綜合以上所述，壓縮感知的技術能促使遠端醫療系統的發展，並且能夠實現輕量化的邊緣運算 (edge computing)，讓智能穿戴式裝置更貼近人類的生活。 然而，壓縮感知雖然能夠克服傳輸頻寬短缺和資料量龐大的問題，它仍會在遠端醫療系統之下遇到以下瓶頸： 第一，壓縮感知降低前端採樣的複雜度，但是若要在後端將壓縮訊號還原，是非常複雜且耗時的； 第二，在資源有限的邊緣裝置 (edge device) 或是現有的壓縮感知還原晶片，不同稀疏度的生理訊號將為大幅影響訊號還原的效率與品質； 第三，生理訊號疾病診斷分析的複雜度取決於輸入訊號的維度，因此若訊號是在原始維度 (original dimension of raw signals) 下進行分析，必須要先將壓縮訊號還原再作分析，如此會面臨耗時的訊號還原與高複雜度疾病診斷的挑戰； 第四，以間歇型心房顫動做為例子，大部分的時間皆為正常竇性心律，異常的竇性心律只會偶而發生，而在現實中，我們比較重視那些有異常的訊號區段，然而現有的壓縮感知還原晶片無法針對特定的訊號區段做還原，只能先將所有訊號還原 (fully reconstruction)，再進行診斷，因此會浪費許多運算資源和能耗在還原那些正常竇性心律或是不感興趣的訊號區段。 為了克服上述挑戰，是有必要提出新穎的系統架構，並探討壓縮感知引擎在遠端醫療應用中的硬體可適性。因此，本論文提出基於訊號拆解的壓縮學習 (compressed learning, CL)，並且提出主動式還原 (on-demand reconstruction, OD-REC)，能夠自行分析有疾病的訊號區段，再單獨針對這些需要被還原的訊號區段，將其從壓縮領域 (compressed domain) 還原回原始的完整訊號。除此之外，為了實現輕量化的遠端醫療系統，本論文將從不同生理訊號之間的稀疏度變化探討其對於壓縮感知應用於遠端醫療的影響: 在面對多變的生理訊號，是否能夠在單位時間內達到最佳的還原品質 (訊號還原效率)或是達到更高的分類準確率。除此之外，我們亦探討生理訊號中稀疏度變化對於壓縮感知引擎的硬體挑戰與可適性，以實現多種不同複雜度之生理訊號還原。最後，我們以 TSMC 40nm CMOS 製程實現一個基於壓縮感知的訊號重現與即時分析的引擎，能夠根據使用者的環境支援三種模式：訊號還原、壓縮領域之訊號分析以及主動式還原。相較於傳統的壓縮感知還原引擎，在面積-能耗效率上能夠有 3.22 倍到 67.57 倍的提升。 綜合以上所述，本論文主要實現基於壓縮感知之低複雜度壓縮領域探勘並分析生理訊號稀疏度對於 CS 後端還原的影響。此外，本論文根據提出的演算法設計一顆即時訊號診斷與還原的晶片，能夠以更低的運算複雜度和更好的面積能耗效率為遠端醫療照護系統提供更多元的功能。

The increasing trends of an aging population and chronic diseases have aroused the awareness of health challenges. Also, due to the trend of healthcare systems moving toward patient-centered, many wearable and wireless biosensors have been invented these years to provide a convenient and comfortable way for monitoring physiological signals. With biosensors and analytics models, physiological signals can be further analyzed by computer-aided diagnostic systems to gather the diagnostic information that is relevant to diseases. Take electrocardiography (ECG) signal monitoring as an example, it is considered as a promising technique to realize healthcare telemonitoring system since ECG signals record from the electrical activity of the heart over a period of time. By analyzing the characteristics of ECG signals, one can significantly improve the quality of life and enable early detection of deterioration. As the wireless wearable physiological sensor nodes are known to resource-limited, such as short battery life and limited bandwidth, it is an important issue to reduce the overall data rate for signal acquisition and enhance the energy efficiency of data transmission. To address these issues, compressed sensing (CS) is an emerging signal processing technique that combines both sampling and compressing with a random sensing matrix. Thus, CS can measure the signals with fewer measurements than the Nyquist rate and recover the signals by CS reconstruction. Since CS needs fewer measurements, it can provide a more efficient way to transmit the compressed data in the wireless healthcare telemonitoring system. Hence, CS is one of the most suitable choices for real-time telemonitoring tasks. For wireless healthcare applications, although CS has successfully reduced the overall transmission data rate and improve power efficiency between sensors and gateways, the wireless healthcare telemonitoring system is still facing the challenge: 1) Insufficient hardware flexibility of CS reconstruction engine for supporting physiological signals with Varied signal sparsity, 2) High complexity CS reconstruction of performing fully reconstruction on every compressed signal, 3) High complexity of analyzing diagnostic information from high-dimensional signals, and 4) Diagnose signals after energy-consuming signal reconstruction. To date, existing CS reconstruction engines can merely support signal reconstruction. Neither do they support signal analytics in the compressed-domain or focus on signals with symptoms of diseases. Hence, this thesis aims to overcome the aforementioned challenges. First of all, we develop a framework that can avoid unnecessary energy-consuming CS reconstruction: compressed analytics with on-demand reconstruction. The proposed framework can classify the received signals in the compressed domain to see if the received signals contain information that is relevant to diseases. Also, with the assistance of on-demand reconstruction, only those high-risk signals are further reconstructed. Furthermore, to achieve low-complexity compressed analytics and CS reconstruction, we investigate the classification performance in the compressed domain and the reconstruction efficiency of CS reconstruction algorithms by considering the sparsity of physiological signals. We show how the signal sparsity impacts the hardware designs of CS reconstruction engines and explore the hardware design space in practical physiological signal monitoring. Finally, based on the proposed framework, we design a tri-mode compressed analytic engine that can support 3 modes: signal reconstruction, classification, and on-demand reconstruction. The proposed engine is fabricated in 40-nm CMOS technology and the measurement results show that the area-energy efficiency (AEE) of this engine is 3.22-to-67.57x better compared with the state-of-art designs. In summary, this thesis focuses on low-complexity CS-based signal analytics while dealing with the physiological signals with varied sparsity. The 2.41-mm2 tri-mode compressed analytic engine provides more comprehensive functionalities but is more lightweight for medical telemonitoring applications.