Developing a hybrid linear model with a multilayer feed-forward neural network for HbA1c modeling among diabetes patients

Abstract

Hemoglobin A1c (HbA1c) is the gold-standard measure for diagnosing and managing diabetes. Given the importance of data-driven decisions, this paper aimed to develop a method for elucidating and predicting HbA1c levels. We developed a comprehensive method for analyzing the multiple linear regression through the R syntax, embedding the multilayer feedforward neural networks (MLFFNN) and bootstrapping. The success of the proposed method was determined by the accuracy of the prediction. The quality of the obtained model was represented by the size of the obtained minimum mean square error (MSE). This study used secondary diabetes data with a total of 1000 observations to illustrate the development method (data obtained after the bootstrapping procedure). The clinical relevance and significance of each preselected variable were evaluated before further testing. The proposed variables were assessed using the MLFFNN methodology, such as the HbA1c, fasting blood sugar (FBS), urea, and blood sodium levels. It was found that FBS, urea, and blood sodium levels can all be used to verify HbA1c. FBS ( = 0.45931; Std SE= 0.01018; p< 0.01), urea ( =-0.03777; SE= 0.00266; p < 0.01), and blood sodium levels ( =-0.06685; SE= 0.01112; p < 0.01) all had a significant impact on HbA1c. Our strategy provides an accurate prediction possible. The methodology precisely assesses the validity of the final model. Superior model performance leads to more efficient management in decision-making.