813-P: Machine Learning–Based Prediction Models for Initial Insulin Pump Dosing in Type 2 Diabetes Patients

Le résumé fourni par la source

Introduction and Objective: Accurate initial insulin dosing is essential for optimal glycemic control in type 2 diabetes patients with insulin pumps. Traditional weight-based estimations lack precision due to the heterogeneity of type 2 diabetes, underscoring the need for advanced predictive approaches. This study developed machine learning models to enhance the accuracy of initial premeal and basal dose predictions. Methods: Data from 1,245 patients at the First Affiliated Hospital of Guangxi Medical University were used for model construction and internal validation, and 60 patients from Sun Yat-sen Memorial Hospital for external validation. Adults aged 18-79 years with type 2 diabetes who initiated insulin pump therapy were included, with data collected during the first 24 hours following admission. Patients with severe comorbidities, acute complications, or organ failure were excluded. A stacked ensemble framework combining random forest, XGBoost, GBM, SVM, and Bayesian regression was used. Model 1 predicts premeal insulin doses, and Model 2 basal doses based on Model 1’s outputs. Performance was evaluated using RMSE, MAE, and MAPE. Results: Model 1 achieved an RMSE of 1.10 IU, MAE of 0.79 IU, and MAPE of 19.10% for internal validation, and an RMSE of 1.21 IU, MAE of 0.88 IU, and MAPE of 17.83% for external validation. Model 2 achieved an RMSE of 2.31 IU, MAE of 1.80 IU, and MAPE of 18.66% for internal validation, and an RMSE of 3.89 IU, MAE of 3.21 IU, and MAPE of 23.47% for external validation. Compared to traditional methods, machine learning models significantly reduced RMSE, MAE, and MAPE in both premeal and basal dose predictions. The prediction models are available as a web-based calculator at https://rongxi.shinyapps.io/Pump/. Conclusion: The machine learning models accurately predict initial insulin pump dosing and outperform traditional methods, offering a practical tool for optimizing therapy in type 2 diabetes patients with insulin pump treatment. Disclosure M. Tang: None. X. Wang: None. X. Rong: None. Funding the Clinical Research 'Climbing' Program of the First Affiliated Hospital of Guangxi Medical University (YYZS2023010); Guangxi Medical University Student Innovation and Entrepreneurship Training Program Project (X202310598348 and S202410598192)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.