Early Diabetes Detection

Abstract

Diabetes, one of the most prevalent metabolic disorders worldwide, requires early and accurate detection for effective management and prevention of complications. This study focuses on addressing the challenges of timely and accurate detection of diabetes through a novel approach that combines machine learning models and different deep learning models. The approach incorporates different models which include Random Forest, Support Vector Machine, and Decision Tree along with different ensemble models. The ensemble models include combinations of different models and classifiers to improve the results. Ensemble techniques like soft voting and hard voting are also used to enhance the results of models and get better accuracy of prediction, which were implemented on three datasets Pima, diabetes dataset and human vital sign dataset. The deep learning models Bidirectional LSTM and Heterogeneous Modified Neural Network were used on the NHANES dataset to achieve improved results. To enable both forward and backward learning, a Bidirectional LSTM model was applied to the NHANES dataset. Additionally, a Heterogeneous Modified Neural Network was implemented to handle the diverse format of data. The proposed work for diabetes detection on PIMA dataset has achieved an accuracy of 91.3% through the implementation of ensemble technique. We also achieved an accuracy of 88% by single classifier, random forest on Pima dataset. For diabetes prediction dataset we have achieved 87% accuracy with single classifier random forest. For NHANES dataset we have achieved 87.5% accuracy with bidirectional LSTM and 85.7% accuracy by heterogeneous modified neural network. The experimental results show that the proposed method achieves better results for diabetes detection.