Diabetes Prediction Using Stacked Ensemble Lstm Model Optimized with Coyote Optimization Algorithm

Abstract

Diabetes mellitus still causes a significant global health concern even if it is a chronic condition defined by elevated blood glucose levels. Inappropriate diabetes mellitus detection or treatment can cause major complications including kidney damage, vision problems, and cardiovascular diseases. Early diagnosis is essential for both lowering these risks and improving patient outcomes; hence, its importance cannot be underlined. Conventional prediction models sometimes run across difficulties related to feature selection and model optimisation that finally produce less than accuracy. This work addresses the issues raised by presenting a Stacked Ensemble LSTM (SE-LSTM) model optimised with the Coyote Optimisation Algorithm (COA) for the aim of diabetes prediction with higher accuracy. The suggested method standardises a large spectrum of attribute scales by means of a robust data preprocessing pipeline. Depending on their interdependence, both the HSIC Lasso approach and Z-score normalisation are part of this pipeline and used to identify most relevant features. The SE-LSTM architecture consists of several LSTM layers to adequately capture temporal dependencies. Conversely, the COA improves hyperparameter tuning by simulating social behaviour of coyotes in their natural environment. With the Pima Indian Diabetes Dataset, the model showed amazing predictive power. Its 98.5% accuracy, 97.8% precision, and 98.2% recall above those of other machine learning models including Random Forest (95.6%) and Gradient Boosting (96.8%). The results show that the SE-LSTM with COA is a good method for diabetes prediction since it provides enhanced generalisation and feature utilisation.