基於多波長PPG和MAML的個人化非侵入式血糖監測儀

Abstract

糖尿病是一種代謝失調的疾病，造成此疾病的主要特徵是長期血糖水平異常升高，如果未能有效管理，將會導致重大的健康危害。儘管目前尚未有治癒方法，但透過嚴密的血糖監控與管理，可以顯著降低長期併發症的發生風險。本研究提出並開發了一種創新輔助型的非侵入式人體血糖連續監測系統，該系統利用多波長光體積描記法(PPG)訊號，LED波長分別為紅外光950 nm、紅光660 nm和綠光525 nm，結合機器學習技術，開發出一組有效的輔助式連續血糖偵測儀器，減少傳統侵入式方法帶來的疼痛和感染風險，同時亦提升了量測便利性。 實驗中，我們的硬體系統是基於MIKROE 2036模組，整合了Arduino Nano，以有效捕捉高品質的PPG信號。為了確保模型預測的準確性，我們使用符合ISO 15197:2013標準的市售血糖儀Accu-Chek Guide作為血糖參考值。透過一系列資料預處理步驟，包括濾波、信號分段及SQI訊號優劣分類器，我們成功提取了高品質的PPG信號數據，並進行後續分析。所提取的特徵涵蓋了生理特徵、血糖相關特徵及統計特徵，這不僅提高了機器學習中模型的可解釋性，也縮短訓練時間並提升了預測精度。 強調了高品質數據對於訓練機器學習模型至關重要。我們從50名健康受試者中收集了數據，受試者平均年齡為24歲，平均BMI為22.7。研究結果顯示，當數據質量超過50%的門檻時，MAML模型顯著優於傳統的簡單人工神經網絡(Simple ANN)模型。MAML模型在測試任務中的平均絕對相對誤差(MARD)為14.98%，相較於簡單ANN模型的24.71%，提升了約40%。除此之外，MAML模型的預測標準差較低，表明其在穩定性上具有更大優勢。通過排除低質量數據(質量指數 < 30)，我們進一步優化了訓練過程，進而提升了模型的整體性能。這些結果經由Clarke誤差網格分析驗證，顯示我們的MAML模型不僅符合臨床標準，且在各種任務中表現出穩定可靠的性能。 總結來說，本論文展示了結合PPG信號與先進機器學習應用的潛力，開發出一種可靠的個人化非侵入式血糖監測系統。這項研究也為未來非侵入式血糖監測創新奠定了堅實的基礎，其最終目標是提升患者的舒適度及健康管理效果。

Diabetes mellitus, a metabolic disorder marked by prolonged elevated blood glucose levels, poses significant health risks if not properly managed. While there is no cure, vigilant monitoring and management of blood glucose levels can mitigate long-term complications. In this study, we proposed and developed an innovative non-invasive blood glucose monitoring device that leverages multi-wavelength Photoplethysmography (PPG) signals at 525 nm, 660 nm, and 950 nm, combined with machine learning techniques. This device provides a convenient alternative to traditional invasive glucose measurement methods, reducing pain, infection risk, and improving usability. Our hardware system, based on the MIKROE 2036 module integrated with an Arduino Nano, effectively captured high-quality PPG signals. The Accu-Chek Guide glucometer, compliant with ISO 15197:2013 standards, was used as a reference for blood glucose values to ensure the accuracy of our model's predictions. Comprehensive data preprocessing, including filtering, signal segmentation, and the application of an SQI classifier, was employed to refine raw PPG signals. This process ensured that only high-quality signals were analyzed. The extracted features—encompassing physiological, glucose-related, and statistical attributes—were fed into machine learning models, enhancing interpretability, reducing training time, and improving prediction accuracy. Our study also emphasized the importance of high-quality data in training machine learning models. We collected data from 50 healthy participants, with an average age of 24 years and an average BMI of 22.7. The findings demonstrated that the Model-Agnostic Meta-Learning (MAML) model significantly outperformed the Simple Artificial Neural Network (ANN) model, especially when data quality exceeded a threshold of 50. The MAML model achieved a Mean Absolute Relative Difference of 14.98% on testing tasks, compared to the Simple ANN’s 24.71%, marking a substantial improvement of approximately 40%. The standard deviation of the MAML model's predictions was also lower, indicating greater stability. By excluding lower-quality data, we optimized the training process, further enhancing the model’s overall performance. These results, validated through Clarke’s Error Grid Analysis, confirmed that our MAML implementation met clinical standards and consistently performed well across various tasks. In conclusion, this study demonstrates the potential of integrating PPG signals with advanced machine learning techniques to develop a reliable, non-invasive blood glucose monitoring system. This work lays the groundwork for future innovations in glucose monitoring, with the goal of improving patient comfort and health outcomes.