ENHANCING DIABETES DIAGNOSIS ACCURACY WITH AN ENSEMBLE LEARNING APPROACH

Abstract

The goal of this thesis is to evaluate how different machine learning algorithms perform in detecting diabetes using the Pima dataset. Three models are examined for their accuracy: Random Forest (72%), Long-Short-Term-Memory Networks (LSTMs) (75.32%), & Extreme-Learning-Machine (ELM) (77.70%). The findings demonstrate that the ensemble model containing ELM, LSTM, and RF outperforms the individual models with an accuracy of 93%. To strengthen the ensemble model further, a unique strategy is proposed, comprising a mix of ELM and LSTM models together with the addition of a Random Forest classifier. Cross-validation is utilised to test the suggested model, indicating its improved performance compared to the individual models, with a mean AUC of 0.93. Subsequently, the model is trained on the complete dataset and achieves an accuracy of 93% and an AUC of 0.89 on the testing set. These results show that the proposed approach efficiently diagnoses diabetes and offers promise for assisting clinical decision-making.