Abstract

Info

Title:: Construction and performance comparison of prediabetes prediction model based on machine learning algorithm

Author(s):: HAO Rui1; MA Kaile1; ZHAO Jingyi1; SUN Xiao1; ZHAO Shuang2; LI Min1; 1. Research Laboratory of Molecular Biology, Guanganmen Hospital, China Academy of Chinese Medical Sciences, Beijing 100053, China 2. College of Traditional Chinese Medicine, Changchun University of Chinese Medicine, Changchun 130117, China

Keywords:: Keywords: prediabetes prediction feature selection optimization algorithm ensemble learning

Abstract:: Abstract: Objective To propose an ensemble learning model integrated with feature selection and automatic hyperparameter optimization mechanism for addressing the problem of high difficulty in identifying prediabetic populations and limited performance of traditional prediction models, so as to improve prediction accuracy and model stability. Methods An ensemble framework was constructed based on Lasso feature selection and optimization algorithm. Lasso eliminated redundant variables in the feature engineering phase, while optimization algorithm was used to automatically tune the key parameters of 4 base models: random forest, support vector machine, extreme gradient boosting, and adaptive enhancement. The model was trained and tested under two setting schemes with the training-test set ratios of 3:1 and 4:1, with accuracy, precision, F1 score, and AUC serving as evaluation metrics. Results Under 3:1 training-test set partitioning scheme, the proposed model showed an AUC up to 0.837, an accuracy of 0.77, and an F1 score of 0.76. With the training-test set ratio adjusted to 4:1, the AUC of the proposed model increased to 0.846, while both accuracy and F1 score reached 0.78. Conclusion The proposed machine learning approach demonstrates strong discriminative capacity for prediabetes prediction, and outperforms the traditional model. This strategy not only improves model performance, but also enhances the level of automation and clinical application potential.