Towards Building a System for Predicting Diabetes and related conditions using Machine Learning

The objective of this paper is to develop a system for predicting diabetes and related conditions in patients using Machine Learning techniques with high degree of accuracy so patients can be treated at an early stage, which could provide a life-saving impact. A Backpropagation Neural Network (BPNN) with 50 nodes in hidden layer and K-Nearest Neighbour (KNN) were created to predict diabetes in patients. A Long Short-Term Memory (LSTM) network and Recurrent Neural Network (RNN) with 100 nodes in hidden layer were created to predict blood glucose levels and generate early warning signs for short-term diabetes complications such as hypoglycaemia, hyperglycaemia and pre-diabetic. The BPNN model achieved the best performance for predicting diabetes with an average classification accuracy of 76.7% and was compared with KNN model which achieved an average classification accuracy of 74.0%. While LSTM model achieved the best performance for predicting blood glucose levels with an average classification accuracy of 90.0%, 88.8% sensitivity, 88.0% specificity, 93.0% positive predictive value and 81.3% negative predictive value, and was compared with RNN model which achieved an average classification accuracy of 84.1%. Obtaining highly accurate predictions on future readings shows potential for the system to be used by healthcare care personnel to determine the right form of treatment at an early stage so patients can be treated in advance. The developed system is at its early stages with two fully working tools and shows promise for further development to increase its effectiveness and performance for complete professional use.