心電圖卷積神經網路分類及頸動脈血流影像分析數據輔助麻醉評估之可行性研究

The Feasibility Study of ECG Classification by Convolutional Neural Network and Carotid Artery Blood Flow Data Analysis by Imaging Processing in Assisting Evaluation of Anesthesia

葉力仁(Li-Ren Yeh)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/電機資訊學院/資訊工程學系/博士(2020年)

摘要

麻醉評估是整個手術圍期中最重要的一環，麻醉科醫師必須就病人整體的病況做詳細的評估。麻醉評估會考慮病人各項狀況：年齡、性別、身高、體重、過往疾病史、現在正在發生的疾病、過敏史等重要資訊，然後對病人的麻醉風險做出分類等級。
　　本篇論文所使用的是美國麻醉醫學會的等級(ASA)，分為五個等級。等級越高，風險越高。一般是用問卷的方式來估計麻醉分類風險(ASA)，然而缺少即時動態ECG及頸動脈超音波血流之評估，因此本研究擬透過心電圖卷積神經網路分類及頸動脈超音波血流影像分析數據輔助麻醉評估之可行性研究。
透過IoT設計能將ECG即時訊號傳送至雲端，再經由卷積神經網路分類成種類型；而模型之建立訓練集則來自MIT-BIH Arrhythmia Database公開資料庫，其波型均利用演算法分割存成二維JPG影像，再由兩位臨床醫師分成4類：QRS Widening (n=21377)、Sinus Rhythm (n=19751)、ST Depression (n=7163)、ST Elevation (n=5899)；CNN採用ResNet、AlexNet、SqueezeNet等三種模型，以訓練集與測試集各為50%；模型效能評估採用準確性及Kappa一致性統計量。
　　再者，透過4位志願者進行動態B-Mode頸動脈超音波造影10秒，存成每秒有30個影格之MPEG檔，再透過影像處理對每一張影格(Frame)估算頸動脈管徑收縮變化，並估算最大與最小管徑之比值，再評估正常人之管徑之比值範圍。
透過MIT-BIH Arrhythmia Database公開資料庫，使用卷積神經網路ResNet、AlexNet、SqueezeNet分類心電圖波型之準確性和Kappa統計量分別為(0.97、0.96）、(0.96、0.95）、(0.75、0.67)。而4位志願者之最大、最小及管徑比值分別為(6.73 mm、5.45 mm、1.24)、(7.03 mm、6.45 mm、1.09)、(6.13 mm、5.16 mm、1.19)和(6.71 mm、6.00 mm、1.12)。
本研究顯示，透過IoT即時量測ECG再經由深度模型辨別4類是可行的，其中以MIT資料庫驗證結果顯示，ResNet之分類準確性及一致性最佳；再者經由分析4位志願者動態B-mode頸動脈變化性之最大與最小管徑範圍約為1.09至1.24之間；本研究建議未來麻醉評估選項應列入IoT即時ECG評估結果及影響頸動脈管徑化比值，以做為參考。
未來仍有很長的一段路必須要走，包括增加更多的受試者並擴增年齡範圍，一者提高CNN模型即時分類效能；二者可以測試不同的深度學習模型；三者可以了解各種心電圖變化對應之頸動脈最大、最小及管徑比值之影嚮。

關鍵字

麻醉評估 ； 心電圖 ； 卷積神經網路 ； 超音波 ； 人工智慧 ； 傅立葉轉換

並列摘要

Anesthesia assessment is the most important part of the entire perioperative period, and the anesthesiologist must make a detailed assessment of the patient's overall condition. Anesthesia assessment considers the patient's various conditions: age, gender, height, weight, past disease history, current disease, allergy history, and other important information to classify the patient's anesthesia risk. The grade used in this paper is the American Society of Anesthesiology (ASA), which is divided into five grades. The higher the level, the higher the risk is. The risk of anesthesia classification (ASA) is generally estimated by questionnaires, but there is a lack of real-time dynamic ECG and carotid artery blood flow assessment. Therefore, this study intends to use an ECG convolutional neural network classification and carotid artery blood flow image analysis data to assist the anesthesia Feasibility study for evaluation.
Through the IoT design, ECG real-time signals can be sent to the cloud, and then classified into various types through the convolutional neural network; and the model is created from the MIT-BIH Arrhythmia Database public database, and the waveforms are stored as two-dimensional JPG images. These images are divided into four categories by two experts: QRS Widening (n=21377), Sinus Rhythm (n=19751), ST Depression (n=7163), ST Elevation (n=5899); The CNN adopts three models such as ResNet, AlexNet, SqueezeNet, etc. The training set and test set are 50% respectively; the accuracy and Kappa consistency statistics are used for the model performance evaluation. In addition, dynamic B-Mode carotid angiography was performed on 4 volunteers for 10 seconds and stored as a MPEG file with 30 frames per second. Then, the image processing was used to estimate the change in carotid artery diameter contraction for each frame. The ratio of the maximum and minimum outer diameters is estimated, and the ratio range of the outer diameters of normal persons is evaluated.
The IoT real-time measurement of ECG mentioned in this research is feasible. Furthermore, the accuracy and Kappa statistics of ResNet, AlexNet, and SqueezeNet are (0.97, 0.96), (0.96, 0.95) and (0.75, 0.67) respectively through the MIT-BIH Arrhythmia Database public database. The maximum, minimum diameters and ratios of the four volunteers were (6.73 mm, 5.45 mm, 1.24), (7.03 mm, 6.45 mm, 1.09), (6.13 mm, 5.16 mm, 1.19), and (6.71 mm, 6.00 mm, 1.12).
This research shows that it is feasible to measure ECG by IoT real-time device and then using a deep learning model to identify 4 categories of ECG. Among them, the best-classified results of the MIT database by ResNet which were shown the highest accuracy and Kappa value. The maximum and minimum diameter range of B-mode carotid artery variability is about 1.09 to 1.24. The suggests in this study that the IoT real-time ECG classified results and carotid artery diameter varying ratio should be included anesthesia evaluation in the future, although those measurements to be an index of anesthesia will be a long road.

並列關鍵字

Pre-anesthesia evaluation ； ECG ； CNN ； Carotid Artery Diameter ； Arduino ； IoT ； AI ； B-mode ultrasound ； FFT

參考文獻

1. Curry J, Harris N. Powering the Environmental Internet of Things. Sensors (Basel). 2019; 19(8): 1940.

2. Chen PH, Cross N. IoT in Radiology: Using Raspberry Pi to Automatically Log Telephone Calls in the Reading Room. J Digit Imaging. 2018; 31(3): 371-378.

3. Fletcher AC, Mura C. Ten quick tips for using a Raspberry Pi. PLoS Comput Biol. 2019; 15(5): e1006959.

4. Chen X, Li H. ArControl: An Arduino-Based Comprehensive Behavioral Platform with Real-Time Performance. Front Behav Neurosci. 2017; 11: 244.

5. Schultz BG, van Vugt FT. Tap Arduino: An Arduino microcontroller for low-latency auditory feedback in sensorimotor synchronization experiments. Behav Res Methods. 2016; 48(4): 1591-1607.