Development of a Cyclic Gradient Boosting Model for Early Prediction of Gestational Diabetes Risk
Keywords:
Gestational Diabetes Mellitus , Academic Success, Time-Series Forecast, SHapley Additive exPlanations , Risk Prediction, Low-Resource Settings, Antenatal Care.Abstract
Gradient Boosting Model is a notable public health challenge in Nigeria, especially in the low resource care facilities lacking adequate facilities for standard GDM diagnosis. This study attempts to develop a machine learning technique capable of accurately interpreting GDM prediction results through the use of clinical data. To accomplish this objective, a Cyclic Gradient Boosting (CGB) algorithm was proposed and analyzed using Shapley Additive exPlanations (SHAP). A total of 918 antenatal patient records were used in this investigation. Based on the results obtained from training and validation process, it was possible to obtain a good performance of the designed CGB model, with 82.6% accuracy and ROC-AUC equal to 0.822 and high specificity (>90%). In addition, it is possible to mention the high precision (62.9%), and the analysis of important features revealed that family history of diabetes (~0.60) and body mass index (BMI=0.42) were the most influential predictors. More detailed investigation of the SHAP values shows that the interaction between heredity and BMI is significant (interaction weight: 0.05), as well as the influence of heredity on weight (0.04). Cyclic boosting allowed obtaining consistent and balanced feature importances. In addition, SHAP helped interpret the model by providing easily comprehensible patient specific explanations. It was found that the combination of predictive power with interpretability is possible using non-invasive routinely collected data. This approach offers a practical solution for early identification of GDM risk in resource-constrained settings as well as support targeted screening and timely intervention.
