運用機器學習發展之個別化血壓偵測演算法

Abstract

量測血壓對於醫生對病患的診斷是個重要的指標，血壓對於我們的生活是息息相關，高血壓會導致心血管疾病，若能及時的發現和治療，能夠有效降低死亡率，現今有許多文獻提出非侵入式連續量測血壓的方式，但大多都不是針對個人的生理狀況去作量測，又或者使用較貴重的儀器ECG去作量測。大部分文獻是使用大量化資料建立血壓推估模型，並不能有效套用在每一位受測者身上，且我們也發現採用非線性的類神經網路效果比線性推估效果來的好，且根據不同參數比較發現使用四種參數(脈搏波傳遞時間、收縮期波峰和舒張期波峰的時間差、舒張期時間和光電容積脈搏波強度比)，再根據不同時間周期更新發現使用四天為更新周期最有效果，本論文提出使用機器學習個別化的血壓偵測演算法，根據每個人的生理狀況建立一個個人化的類神經網路模型，採用非線性的方式推估血壓更能有效的偵測血壓，我們使用兩個PPG感測器來獲取所需的生理參數，並且隨時更新所使用的類神經網路訓練資料。本論文針對三位受測者進行長期的血壓量測，並套用本論文的演算法皆得到良好的推估血壓效果，從參考文獻所推估的舒張壓和收縮壓誤差(9.64%、8.06%和14.82%、14.62%)透過本論文演算法下降到4.68%和2.93%且沒有一天的血壓誤差值超過10%。

Blood pressure is an important physiological measure because hypertension can lead to cardiovascular diseases. If we can detect it and treat on time, we can effectively reduce the mortality rate. There are many papers that propose non-invasive and continuous measurement of blood pressure. However, most of them do not measure the physiological condition for each individual, or opt to use the more expensive instrument ECG for measurement. Most of the papers use large amounts of data to establish blood pressure estimation model, but they cannot be effectively applied to each person who undergoes this test. Moreover, we also found that using the nonlinear neural network effect is better than linear estimation. According to the comparison of different parameters, four parameters (pulse wave transit time, systolic peak and diastolic peak time, diastolic time and photoplethysmographic intensity ratio) were discovered, and then we would update them accordingly for different time periods and found that a four-day update cycle is the most effective. This paper proposes to use machine learning customized blood pressure detection algorithm to build a customized neural network model which is based on each person's physiological condition, using non-linear ways to estimate and more effectively detect blood pressure. We use Two PPG sensors to obtain the required physiological parameters and to update the training data of the neural network immediately. In this article, long-term blood pressure measurement of three subjects and applying the algorithm in this paper effectively provide a good estimate of blood pressure. Compared to the estimated diastolic and systolic blood pressure errors reported in literature (9.64%, 8.06% and 14.82%, 14.62%), the errors dropped to 4.68% and 2.93% through the algorithm of this dissertation and none of them had the blood pressure error value of over 10%.