利用CUDA平行化的經驗模態分解應用於心電圖推導呼吸訊號

Abstract

心電圖推導呼吸訊號（EDR）是一種用來從心電圖（ECG）中推導呼吸訊號的技術，它能夠克服傳統擷取呼吸方法的限制。經驗模態分解（EMD）是一個自適應的分析過程，它能夠應用在非線性與非穩態的資料上例如：心電圖，因此能夠用來處理EDR應用。EMD是藉由迭代地將資料分解成數個本質模態函數來分析。傳統上，EMD是序列的運算每一個資料點，因此造成它的執行時間與資料量呈正比關係。在這篇論文中，一個由CUDA語言實作且運作在通用圖形處理器(GPGPU)的平行化EMD演算法被提出，用以改善傳統EMD的效能。並且額外的合併式三次樣條插值與GPU加速技巧被加入用來達到高平行度與高精確度。資料庫測試顯示，我們的CUDA平行化EMD在一百萬個資料點數下達到6.6倍加速，且經過了50次的迭代後只有0.0003%誤差。對於EDR應用，我們的平行化EMD達到69.75%的準確率，並且只需要7.91秒即可處理一分鐘的ECG。

ECG-Derived Respiratory (EDR) is a technique to derive respiratory from electrocardiography (ECG), which can help to overcome limitation of traditional respiratory acquisition method. Empirical Mode Decomposition (EMD) is process of adaptive analysis applicable to non-linear and non-stationary data such as ECG, hence it can be used to deal with EDR application. EMD analyzes data by iteratively decomposing data into multiple Intrinsic Mode Functions (IMFs). Traditionally, EMD is computed on all data points in a serial manner, thus making its execution time grows linearly with the data size. In this work, a parallelized EMD algorithm working on a General-Purpose computing on Graphics Processing Units (GPGPU) in CUDA language is proposed to improve performance over traditional EMD. Moreover, additional merging cubic spline interpolation and GPU acceleration techniques are also incorporated for achieving high parallelism and high accuracy. Statistical result of database shows that our parallelized EMD in CUDA achieves 6.6X speedup with 0.0003% error after 50 times iteration on datasets of 1-million points. For EDR application, our parallelized EMD achieves average 69.75% accuracy with average execution time of 7.91 second for 1-minute windows ECG from Fantasia Database.