A data-driven machine learning algorithm to predict the effectiveness of inulin intervention against type II diabetes.

INTRODUCTION

The incidence of type 2 diabetes mellitus (T2DM) has increased in recent years. Alongside traditional pharmacological treatments, nutritional therapy has emerged as a crucial aspect of T2DM management. Inulin, a fructan-type soluble fiber that promotes the growth of probiotic species like and , is commonly used in nutritional interventions for T2DM. However, it remains unclear which type of T2DM patients are suitable for inulin intervention. The aim of this study was to predict the effectiveness of inulin treatment for T2DM using a machine learning model.

METHODS

Original data were obtained from a previous study. After screening T2DM patients, feature election was conducted using LASSO regression, and a machine learning model was developed using XGBoost. The model's performance was evaluated based on accuracy, specificity, positive predictive value, negative predictive value and further analyzed using receiver operating curves, calibration curves, and decision curves.

RESULTS

Out of the 758 T2DM patients included, 477 had their glycated hemoglobin (HbA1c) levels reduced to less than 6.5% after inulin intervention, resulting in an incidence rate of 62.93%. LASSO regression identified six key factors in patients prior to inulin treatment. The SHAP values for interpretation ranked the characteristic variables in descending order of importance: HbA1c, difference between fasting and 2 h-postprandial glucose levels, fasting blood glucose, high-density lipoprotein, age, and body mass index. The XGBoost prediction model demonstrated a training set accuracy of 0.819, specificity of 0.913, positive predictive value of 0.818, and negative predictive value of 0.820. The testing set showed an accuracy of 0.709, specificity of 0.909, positive predictive value of 0.705, and negative predictive value of 0.710.

CONCLUSION

The XGBoost-SHAP framework for predicting the impact of inulin intervention in T2DM treatment proves to be effective. It allows for the comparison of prediction effect based on different features of an individual, assessment of prediction abilities for different individuals given their features, and establishes a connection between machine learning and nutritional intervention in T2DM treatment. This offers valuable insights for researchers in this field.