以機器學習法建立糖尿病營養衛教門診病人血糖變化之預測模型與系統以輔助臨床決策

Using Machine Learning Algorithms to Build a Prediction Model and System for Blood Sugar Change that Assists Nutritional Decision-Making for Diabetes Mellitus

目的：以機器學習技術建立有效預測糖尿病個案之長期血糖變化模型，並實作成系統，提供予病人營養介入輔助決策之參考。

材料與方法：數據來源為奇美醫療體系三院區門診營養衛教系統2007至2019年加入糖尿病試辦計畫之成年病人就診營養衛教紀錄，依文獻與專業經驗選擇20個特徵變數，以多種機器學習演算法建立「預測一年後病人糖化血色素是否改善達7%以上」之模型。最後挑選最佳模型（Area Under the Curve[AUC]最高者）實作成預測系統以供臨床使用。

結果：各機器學習法建立之模型精確度在0.735~0.749間，其中支持向量機法之敏感性達0.757、特異性0.739、AUC值0.828，為最佳模型。我們將預測系統提供給3位營養師試用，均獲得正面的肯定，認為此系統對糖尿病營養諮詢衛教非常有幫助。

結論：以機器學習建立之預測模型具有優異的品質，為糖尿病營養諮詢衛教提供非常有前景的方法，可作為臨床疾病照護及飲食衛教介入之有效參考，使病人維持良好之長期血糖控制，減少糖尿病引發合併症發生率，有助於提升醫療品質與促進醫病共享決策。

 

Purpose: The prevalence of diabetes mellitus (DM) continues to increase worldwide. We built a machine learning model and developed a prediction system that is based on an optimal model to effectively predict blood sugar changes in patients with diabetes. Our findings contribute to the implementation of long-term patient nutrition interventions.

Method: Data of outpatients with type 2 DM who were 20 years or older and underwent nutrition education under a diabetes pay-for-performance program were obtained from the Nutrition and Health System Database of the outpatient clinic of the Chi Mei Hospital network; the data spanned the years from 2007 to 2019. On the basis of literature findings and professional experience, 20 characteristic variables and multiple machine learning algorithms were applied to build a model to predict whether the glycosylated hemoglobin (HbA1c) of the outpatients improved by more than 7% after 1 year. The optimal model (model with the highest area under the curve [AUC]) was selected and used to develop a prediction system for use in clinical settings.

Results: The accuracy levels of the developed models ranged from 0.735 to 0.749; the supportvector- machine model with a sensitivity of 0.757, a specificity of 0.739, and an AUC of 0.828 was the optimal prediction model. The prediction system was tested by three dietitians, who affirmed its usefulness for diabetes meal planning and patient health education.

Conclusion: The prediction model based on machine learning algorithms performed excellently, and it is a promising tool for diabetes meal planning and patient health education. It is also an effective supporting tool for clinical disease care and dietary health education interventions. We believe that the model can help patients maintain favorable long-term blood sugar control, reduce their incidence of diabetes-related complications, improve the quality of medical care and promote shared decisionmaking.