Development and validation of a GBDT-based for detection of diabetes

Abstract

Improved early detection and intervention can lessen the substantial health and financial burden of type 2 diabetes (T2D). In order to achieve early prediction of diabetes, an approach for diabetes prediction based on Gradient Boosting Decision Tree (GBDT) was proposed in the present study. The feature selection approach based on GBDT was designed to identify the most relevant features, where GBDT builds a powerful model by combining multiple weak classifiers, which can reduce the risk of overfitting and improve the model generalization ability. Condition prediction was performed based on GBDT and the filtered features, and the truncation value of the feature was calculated. The experimental results in the paper showed that the Area Under the Curve (AUC) value of GBDT was 0.9788, which was a big improvement compared with other studies; the AUC value based on glycated hemoglobin level was 0.7307, and the cutoff value of glycated hemoglobin level was about 6.8, which was very accurate. The prediction of diabetes based on GBDT can help patients to understand whether they have diabetes initially based on their own glycated hemoglobin values, and it can also help clinicians to make more objective judgments in clinical diagnosis in order to judge the patient's situation and subsequent monitoring of the condition, making an excellent contribution to the control of the patient's condition.