Machine Learning for Glucose Prediction to Identify Diabetes-related Metabolic Pathways

Anonymous

Keywords: Neural Network, Gene Expression, Diabetes, Glucose, Biological Pathways

Abstract: Diabetes is a major global health issue, with cases predicted to rise from 451 million today to 642 million by 2040. We use three machine learning methods (k-NN, regression and neural networks) to predict the glucose levels of mice based on genetic expression data across five tissues. Based on the best-performing neural network model, we derive modules that correspond to metabolic pathways by retraining networks on permuted feature clusters. The neural networks performed the best of the three model families, achieving a mean absolute percentage error of 26.0\% on the adipose tissue. From the neural network models, we produce a list of 9 modules that have high impact in their respective models.