基於機器學習與多波長PPG訊號之非侵入式血糖估計

Abstract

根據國際糖尿病聯盟(IDF)的統計，全球有約5.37億人口患有糖尿病，且患病人口預估在2030年及2045年分別達到6.43億及7.83億，而台灣的糖尿病盛行率又高於全球平均，平均十人之中就有一人患病，且據估計全台已患病卻未診斷的人數約占全台人口的1%。通常從罹患糖尿病開始，各種慢性病也會隨之併發，然而糖尿病初期通常沒有明顯症狀，所以對其的預防及檢測不容忽視。血糖的量測不管是對糖尿病患者的健康監控或對未患病者的預防檢測都非常重要，然而現行的量測方法大多是採用侵入式量測或高成本的儀器，這使得對血糖的日常監控變得不方便，因此許多團隊開始研究非侵入式且低成本的血糖量測方法，隨著穿戴式裝置的發展，近年透過光體積變化描記圖(PPG)量測血糖的方法開始被廣泛研究，透過對量測光訊號並提取出相關特徵，並搭配機器學習對血糖做出預測，相關的技術日益成熟中。 本研究提出了一個可量測紅光、紅外光及綠光等三波長的反射式PPG量測系統，透過濾波、訊號切割及一個基於支持向量機(SVM)的訊號品質分類器進行訊號處理，接著從訊號中萃取出與血糖相關之特徵並訓練隨機森林及XGBoost等機器學習模型對血糖進行回歸預測，藉此實現非侵入式且低成本的血糖估算。研究中使用了來自10名受試者的訓練資料，對模型進行訓練之後，以平均絕對相對差(MARD)及Clark’s error grid對模型進行評估，根據評估的結果再優化模型參數並分析各個特徵對於血糖估計的重要性及彼此的相關性。

According to statistics from the International Diabetes Federation (IDF), approximately 537 million people worldwide suffer from diabetes. This number is projected to rise to 643 million by 2030 and 783 million by 2045. Taiwan''s diabetes prevalence rate surpasses the global average, with an average of one in ten individuals diagnosed with the disease. It is estimated that around 1% of Taiwan''s population has undiagnosed diabetes. Once diagnosed with diabetes, patients often subsequently develop other chronic diseases. However, initial stages of diabetes usually lack overt symptoms, emphasizing the importance of its prevention and detection. Monitoring blood sugar levels is vital, both for those diagnosed with diabetes and for prevention in non-diabetic individuals. Contemporary measurement techniques often require invasive procedures or employ high-cost instruments, making routine monitoring challenging. As a result, many teams have begun researching non-invasive and cost-effective blood glucose measurement methods. With the rise of wearable technology, the utilization of Photoplethysmography (PPG) for estimating glucose levels has garnered significant attention. By analyzing the optical signals to extract relevant features and harnessing machine learning for glucose predictions, this technique is showing promising advancements. This study introduces a reflective PPG system capable of detecting red, infrared, and green wavelengths. The system processes signals through filtering, segmentation, and a Signal Quality Index based on the Support Vector Machine (SVM). Subsequently, features related to blood sugar are extracted from these signals and used to train machine learning models like Random Forest and XGBoost for regression predictions on blood glucose levels. This aims to realize a non-invasive and cost-effective blood glucose estimation. After training models using data from ten participants, the models were evaluated based on the Mean Absolute Relative Difference (MARD) and Clark’s error grid. Based on these evaluation results, the hyperparameters were further optimized, and the importance of each feature was analyzed, along with the correlations between them.