以腕戴式裝置監測生理資訊於心臟衰竭病患建立預測再住院率之機器學習模型

Uses Wrist-worn Devices to Monitor Physiological Information in Heart Failure Patients to Predict the Readmission Rate By A Machine Learning Model

賴俞雯(Yu-Wen Lai)

指導教授 ： 林永松

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202203670

摘要

慢性心臟衰竭是一種臨床症候群，患者的再住院和死亡率非常高，預後情形不佳。光體積變化描記圖法(PPG)與穿戴式裝置之結合使實時且無創蒐集患者出院後的生理資訊成為可能，及時預警可能發生的不良事件。
在本論文中，我們針對40名病患PPG與3軸加速度信號紀錄進行再住院風險分析。將信號進行去噪後，與病患性別、年齡、心衰等級結合為模型輸入向量，用於心衰患者的再住院風險分析。實驗依據對加速度數據的應用分成三種方法，方法1將加速度作為濾波器，當加速度值超過閾值時，刪除同時段的PPG。方法2和方法3的加速度數據不作為濾波器，而是提取加速度統計特徵改進模型決策空間。方法2是方法3的對照組，使用文獻中常用的時域統計特徵；方法3統計超過閾值的採樣點，稱為活動力特徵，為本文所提出。接著結合PPG、加速度和病患統計特徵為輸入向量，以LightGBM、XGBoost和隨機森林作為分析模型，選擇其中表現最好的模型進行10-Fold交叉驗證。
本文的貢獻有兩個。首先，以便攜式裝置蒐集的生理信號數據為主體，建立心衰病患一年再入院的風險分析模型是一項較創新的挑戰。本文的實驗在測試集上最佳結果有96.67%的準確率與0.99的AUC，並在以病患為主體分群的10-Fold交叉驗證中取得84.58%的平均準確率和0.90的平均AUC。第二是本文所提出之活動力特徵，經由實驗證明對模型判斷有高度效益。相較於過濾活動力大PPG區段的方法，提升14%；相較於選用一般統計特徵的對照組方法在10-Fold交叉驗證中提高5%。

關鍵字

光體積變化描記圖法 ； 心臟衰竭 ； 信號處理 ； 3軸加速度 ； 心率變異分析 ； 再住院

並列摘要

Chronic heart failure is a clinical syndrome, the patient's rehospitalization and mortality are very high, and the prognosis is poor. The combination of Photoplethysmography (PPG) and wearable devices makes it possible to collect the physiological information of patients after discharge in real-time and non-invasively, and to provide timely warning of possible adverse events.
In this thesis, we performed a readmission risk analysis for 40 patients with PPG and 3-axis acceleration signal records. After denoising the signal, it is combined with the patient's gender, age, and heart failure grade as the model input vector, which is used for the risk analysis of rehospitalization of heart failure patients. The experiment is divided into three methods based on the application of acceleration data. In Method-1, the acceleration is used as a filter. When the acceleration value exceeds the threshold, the PPG of the same period is deleted. The acceleration data of Method-2 and Method-3 are not used as a filter, but the statistical acceleration features are extracted to improve the model decision space. Method-2 uses the time-domain statistical features commonly used in the literature and is the control group of Method-3; Method-3 counts the sampling points that exceed the threshold, called activity features, which are proposed in this thesis. Then combined with PPG, acceleration, and patient statistical features as input vectors, LightGBM, XGBoost, and random forest were used as analysis models, and the best performing model was selected for 10-Fold cross-validation.
There are two contributions of this thesis. First, it is a relatively innovative challenge to establish a risk analysis model for one-year readmission of heart failure patients based on the physiological signal data collected by portable devices. The best results on the test dataset in this thesis have an accuracy of 96.67% and an AUC of 0.99, and achieved an average accuracy of 84.58% and an average AUC of 0.90 in the 10-Fold cross validation with patients as the main group. The second is the activity feature proposed in this thesis, which has been proved to be highly effective for model judgment through experiments. Compared with the method of filtering the PPG section with high activity, the improvement is 14%; compared with the control group experiment using general statistical features, the improvement in 10-Fold cross validation is 5%.

並列關鍵字

Photoplethysmography ； Heart Failure ； Signal Processing ； 3-axis Acceleration ； Heart Rate Variability ； readmission

參考文獻

[1] C. W. Tsao, A. W. Aday, Z. I. Almarzooq, A. Alonso, A. Z. Beaton, M. S. Bittencourt, A. K. Boehme, A. E. Buxton, A. P. Carson, Y. Commodore-Mensah, et al.,“Heart disease and stroke statistics—2022 update: a report from the american heart association,” Circulation, vol. 145, no. 8, pp. e153–e639, 2022.

[2] F. Sarijaloo, J. Park, X. Zhong, and A. Wokhlu, “Predicting 90 day acute heart failure readmission and death using machine learning-supported decision analysis,” Clinical cardiology, vol. 44, no. 2, pp. 230–237, 2021.

[3] 衛生福利部統計處, “110 年死因統計結果分析 (含附表),” 2022.

[4] R. Coronel, J. De Groot, and J. Van Lieshout, “Defining heart failure,” Cardiovascular research, vol. 50, no. 3, pp. 419–422, 2001.

[5] R. C. of Physicians, Chronic Heart Failure: National Clinical Guideline for Diagnosis and Management in Primary and Secondary Care. UK: National Clinical Guideline Centre, 2010.