Scholarlink

Background <p>Glucose control of ICU patients is complicated due to various diseases, different patient characteristics, and divergent treatment received during ICU stay. To deal with this challenge, machine learning prediction models are emerging to forecast glucose-related events. Our study utilizing the machine learning algorithm to develop a real-time blood glucose (BG) prediction model, which aims to promote the precision insulin therapy.</p> Methods <p>Electronic medical record data from a tertiary hospital were collected from September 2021 to September 2022. Demographics, comorbidities, vital signs, laboratory results, previous BG measurements, records of insulin injection, nutritional intake, and vasoactive inotropic score (VIS) were included as candidate predictor variables. Four machine learning algorithms including elastic net linear regression, random forest, eXtreme Gradient Boosting (XGBoost), and support vector regression were employed to develop prediction model. Internal validation was performed to select final model, and then was externally validated using MIMIC-IV database. Performance metrics included root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute deviation (MAD), and Clarke error grid analysis (CEGA). Finally, subgroup analysis detected predictive performance’s heterogeneity, while Shapley additive explanation (SHAP) method enhanced the clinical interpretability.</p> Results <p>Total of 3,385 ICU patients with 58,090 BG measurements were included, with 80% for training and 20% for validation. XGBoost performed best in internal validation and was selected as final model (RMSE = 33.59mg/dL, MAPE = 15.39%, MAD = 23.92mg/dL, proportion of type A predictions = 75.67%, and proportion of clinically inappropriate predictions = 0.29%). XGBoost model also showed satisfactory performance in external validation of MIMIC-IV database, underscoring robustness and generalizability. Finally, SHAP method elucidated predictors’ contributions to the predictions from both global and individual perspectives.</p> Conclusions <p>Our study proposes a real-time blood glucose prediction model utilizing the XGBoost machine learning framework in real-world ICU scenarios. The model could facilitate clinicians in understanding the real-time trend of patients’ glucose level and making timely therapeutic decisions.</p> Clinical trial number <p>Not applicable.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development and validation of a machine learning model for real-time blood glucose prediction for ICU patients

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月9日

查看全文
  • Shining Cai,
  • Yundi Hu,
  • Yixiang Hong,
  • Luheng Qian,
  • Shilong Lin,
  • Xiaolei Lin,
  • Ming Zhong,
  • Yuxia Zhang

摘要

Background

Glucose control of ICU patients is complicated due to various diseases, different patient characteristics, and divergent treatment received during ICU stay. To deal with this challenge, machine learning prediction models are emerging to forecast glucose-related events. Our study utilizing the machine learning algorithm to develop a real-time blood glucose (BG) prediction model, which aims to promote the precision insulin therapy.

Methods

Electronic medical record data from a tertiary hospital were collected from September 2021 to September 2022. Demographics, comorbidities, vital signs, laboratory results, previous BG measurements, records of insulin injection, nutritional intake, and vasoactive inotropic score (VIS) were included as candidate predictor variables. Four machine learning algorithms including elastic net linear regression, random forest, eXtreme Gradient Boosting (XGBoost), and support vector regression were employed to develop prediction model. Internal validation was performed to select final model, and then was externally validated using MIMIC-IV database. Performance metrics included root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute deviation (MAD), and Clarke error grid analysis (CEGA). Finally, subgroup analysis detected predictive performance’s heterogeneity, while Shapley additive explanation (SHAP) method enhanced the clinical interpretability.

Results

Total of 3,385 ICU patients with 58,090 BG measurements were included, with 80% for training and 20% for validation. XGBoost performed best in internal validation and was selected as final model (RMSE = 33.59mg/dL, MAPE = 15.39%, MAD = 23.92mg/dL, proportion of type A predictions = 75.67%, and proportion of clinically inappropriate predictions = 0.29%). XGBoost model also showed satisfactory performance in external validation of MIMIC-IV database, underscoring robustness and generalizability. Finally, SHAP method elucidated predictors’ contributions to the predictions from both global and individual perspectives.

Conclusions

Our study proposes a real-time blood glucose prediction model utilizing the XGBoost machine learning framework in real-world ICU scenarios. The model could facilitate clinicians in understanding the real-time trend of patients’ glucose level and making timely therapeutic decisions.

Clinical trial number

Not applicable.