A Pipeline-based Framework for Early Prediction of Diabetes

Diabetes is a chronic disease worldwide, with an increasing annual death rate. Many health professionals seek innovative ways to detect and treat it early. Rapid advances in machine learning have improved disease diagnosis. However, because of the small amount of labeled data, the frequency of null and missing values, and the imbalance of databases, creating an optimal predictor for disease diagnosis has become a great challenge. This study aimed to present a pipeline-based classification framework for predicting diabetes on two datasets of Indian diabetic patients with two classes (patient and healthy groups) and Iraqi with three classes (patient, healthy, and prediabetes groups).

An important part of this framework is preprocessing. Different ML models based on the One-Vs-One approach for the three-class mode are implemented in the framework. Because of the imbalance of the data set, besides the accuracy evaluation criterion, the area under the receiver operating characteristic (ROC) curve is also used. To increase the level of these two criteria, the Hyper-parameters of each model are optimized using optimization methods to build a powerful model with less training and testing time through various feature selection methods.

The proposed framework was assessed for diabetes prediction on two datasets of Indian and Iraqi diabetic patients. It was revealed that using AdaBoost for the Indian dataset (ACC=89.98, AUC=94.11) and random forest for the Iraqi dataset (ACC=98.66, AUC= 98.62), good accuracy and performance were obtained.

Regarding ACC parameters, precision, accuracy, recall, and F1-Score, the pipeline-based framework has an optimal performance in predicting diabetes, therefore, it can be used in clinical decision support systems.