近年來各式不同的糖尿病管理輔助工具陸續被發明，如連續型血糖監測(Continuous Glucose Monitoring, CGM)、胰島素幫浦(Insulin Pump)等，但由於技術的尚未成熟到足以完全投入使用，又或者是高昂的價格等因素，使得傳統血糖監測(Blood Glucose Monitoring, BGM) 仍是大多數的第一型糖尿病(Type 1 Diabetes, T1D) 病人的主要使用工具。這項研究旨在不依靠新型的糖尿病管理工具，針對僅使用BGM 的T1D病人提供一套完整的胰島素劑量最佳化模型。研究強調了T1D病人面臨的挑戰，包括低血糖風險、不完善的血糖管理導致的併發症以及目前CGM技術的侷限性，包括了其高成本和校準要求限制了其在全球第一型糖尿病病人間的可及性。本研究使用REPLACE-BG資料集結合先進的機器學習模型，為主要依賴BGM的T1D病人提供最佳化的胰島素劑量建議。並透過隨機森林模型(Random Forest Model) 達到RMSE 0.58 U的成果，顯著地提升了機器學習應用於胰島素最佳化的表現結果。透過將機器學習與糖尿病管理相結合，本研究成功幫助病人最佳化胰島素劑量，進而使病人能夠更好地進行糖尿病管理、提高生活品質，同時由於本研究著重於BGM設備，也提供病人相較CGM設備更加符合成本效益的替代方案。;In recent years, various diabetes management tools like Continuous Glucose Monitoring (CGM) and Insulin Pumps have been invented. However, due to technological immaturity or high costs, traditional Blood Glucose Monitoring (BGM) remains the primary tool for most Type 1 Diabetes (T1D) patients. This study aims to develop a comprehensive insulin bolus optimization model for T1D patients using only BGM. It highlights challenges such as hypoglycemia risks, complications from inadequate glucose management, and CGM limitations. Despite CGM advancements, its high cost and calibration requirements limit accessibility. Using the REPLACE-BG dataset combined with advanced machine learning models, the study provides optimized insulin bolus recommendations for T1D patients primarily relying on BGM. Achieving RMSE 0.58 U with the Random Forest model, the study significantly enhances insulin optimization. By integrating machine learning with diabetes management, this study successfully optimized insulin bolus for patients, helping them better manage their diabetes and improve their quality of life. Additionally, as the study focuses on BGM devices, it provides a more cost-effective alternative for patients than CGM devices.
