PREDICTION OF TYPE2 DIABETES USING ARTIFICIAL INTELLIGENCE METHODS

Abstract

This paper explores the usefulness of incorporating a feature based on fuzzy logic with machine learning models in predicting diabetes. Diabetes is a multifactorial and complex disease where physiological variables interrelate with each other and traditional machine learning methods fail to represent such non-linear interrelations and uncertainty of clinical data. In order to overcome this shortcoming, a fuzzy inference system was created to produce a composite risk feature (denoted as fuzzy_risk) that captures domain knowledge and reflects the overall impact of critical clinical indicators. The three classification models that were trained and assessed were Decision Tree (DT), K-Nearest Neighbors (KNN) and Multi-Layer Perceptron (MLP) neural network in the presence of the fuzzy-derived feature and the absence of this feature. Preprocessing of the dataset was conducted following standard methods, such as stratified data division and normalization of features, in order to compare the models fairly. Accuracy, weighted F1-score and area under the receiver operating characteristic curve (AUC-ROC) were used to measure model performance, particularly the effect of class imbalance. The findings indicate that the fuzzy_risk feature is able to yield similar positive performance in all the models. Nevertheless, the extent of improvement is small, and it is usually between 1 and 2 per cent in the measures of evaluation. The results also indicate that the simpler models like Decision Tree and KNN can take advantage of the fuzzy feature as opposed to the neural network, which can learn non-linear relationships on its own. Also, all the models merge to the same performance range which shows the existence of a performance ceiling imposed by the data. In practical terms, fuzzy feature increases the interpretability by giving us a structured risk representation consistent with clinical reasoning. This enhances transparency and can be adopted in health care environments. The study, in general, shows a significant but gradual effect of incorporating the fuzzy logic with machine learning on diabetes prediction, noting the role of feature engineering and data quality in the development of even greater improvements.