Özet:

Diabetes is a common disease, incurable and fatal in its complication phases. Its management, like many other metabolic diseases, remains a scientific challenge. Mathematical approaches have been used to understand this scourge and artificial intelligence is used to model its prediction. In general, the effectiveness and efficiency of an artificial intelligence solution depends on the nature and characteristics of the data and the performance of the learning methods. Hence the interest in the quality of the data and the performance of the methods used to model such a task. In order to find a suitable artificial intelligence model for diabetes prediction, several studies have used methods from different techniques. Thus, diabetes prediction has been addressed using machine learning methods, neural networks, deep learning, Bayesian naive classification, K-nearest neighbors and machine vector support. In order to compare the performance to determine the best model, several of these methods are analyzed in previous studies. Thus, this paper evaluates the methods based on the decision tree technique (DT, RF, LightGBM, Adaboost and XGBoost), based on the PIMA Diabetes Indian data (PID). The aim is to show the predictive ability of the methods of this technique and to determine the appropriate method for predicting diabetes with raw data. The PIMA data are described statistically, and the comparative analysis of the models is performed following K-fold cross-validation, before and after class balancing. At the end of the experiment, the best results are obtained by LightGBM, XGBoostand RF on different metrics.

Anahtar Kelimeler: