Effect of dimension reduction with PCA and machine learning algorithms on diabetes diagnosis performance

Abstract

Diabetes, a long-term metabolic disorder, causes persistently high blood sugar and presents a significant global health challenge. Early diagnosis is of vital importance in mitigating the effects of diabetes. This study aims to investigate diabetes diagnosis and risk prediction using a comprehensive diabetes dataset created in 2023. The dataset contains clinical and anthropometric data of patients. Data simplification was successfully applied to clean unnecessary information and reduce data dimensionality. Additionally, methods like Principal Component Analysis were applied to decrease the number of variables in the dataset. These analyses rendered the dataset more manageable and improved its performance. In this study, a dataset encompassing health data of a total of 100,000 individuals was utilized. This dataset consists of 8 input features and 1 output feature. The primary objective is to determine the algorithm that exhibits the best performance for diabetes diagnosis. There was no missing data during the data preprocessing stage, and the necessary transformations were carried out successfully. Nine different machine learning algorithms were applied to the dataset in this study. Each algorithm employed various modelling approaches to evaluate its performance in diagnosing diabetes. The results demonstrate that machine learning models are successful in predicting the presence of diabetes and the risk of developing it in healthy individuals. Particularly, the random forest model provided superior results across all performance metrics. This study provides significant findings that can shed light on future research in diabetes diagnosis and risk prediction. Dimensionality reduction techniques have proven to be valuable in data analysis and have highlighted the potential to facilitate diabetes diagnosis, thereby enhancing the quality of life for patients.