Two Machine-learning Hybrid Models for Predicting Type 2 Diabetes Mellitus

Abstract

Background: The global increase in diabetes prevalence necessitates advanced diagnostic methods. Machine learning has shown promise in disease diagnosis, including diabetes.

Materials and methods: We used a dataset collected from the Medical City Hospital laboratory and the Specialized Center for Endocrinology and Diabetes at Al-Kindy Teaching Hospital in Iraq. This dataset includes 1000 physical examination samples from both male and female patients. The samples are categorized into three classes: diabetic (Y), nondiabetic (N), and predicted diabetic (P). The dataset contains twelve attributes and includes outlier data. Outliers in medical studies can result from unusual disease attributes. Therefore, consulting with a specialist physician to identify and handle these outliers using statistical methods is necessary. The main contribution of this study is the proposal of two hybrid models for diabetes diagnosis in two scenarios: (1) Scenario 1 (presence of outlier data): Hybrid Model 1 combines the K-medoids clustering algorithm with a Gaussian naive Bayes (GNB) classifier based on kernel density estimation (KDE) to handle outliers and (2) Scenario 2 (after removing outlier data): Hybrid Model 2 combines the K-means clustering algorithm with a GNB classifier based on KDE with suitable bandwidth. We performed principal component analysis to minimize dimensionality and evaluated the models using fivefold cross-validation.

Results: All experiments were conducted in identical settings. Our proposed hybrid models demonstrated superior performance in two scenarios, handling and rejecting outliers, compared to other machine-learning models in this study, including support vector machines (with radial-based, polynomial, linear, and sigmoid kernel functions), decision trees (J48), and GNB classifiers for diabetes prediction. The average accuracy for Scenario 1 with Hybrid Model 1 was 0.9743, and for Scenario 2 with Hybrid Model 2, it was 0.9867. We also evaluated precision, sensitivity, and F1-score as performance metrics.

Conclusion: This study presents two hybrid models for diabetes diagnosis, demonstrating high accuracy in distinguishing between diabetic and nondiabetic patients and effectively handling outliers. The findings highlight the potential of machine-learning techniques for improving the early diagnosis and treatment of diabetes.

Keywords: Decision tree; Gaussian naive Bayes; K-means; K-medoids; diabetes mellitus prediction; kernel density estimation; support vector machine.

Figure 1

(a). An example of a classification problem of 3 classes: green, red, and blue. (b). The red-blue line is generated via a one-to-one technique to maximize the distance between the blue and red spots. Green points have nothing to do with it

Figure 2

The Iraqi Patient Dataset for Diabetes dataset population distribution of all attributes, where the green, blue, and yellow color distributions indicate diabetic (C0) individuals, nondiabetic (C1) individuals, and predicted diabetic (C2) individuals, respectively

Figure 3

M1. In the first scenario, which involves the presence of outliers. Thus, we used K-medoids clustering methods and then naive Bayes-based Kernel density estimation

Figure 4

M2. In the second scenario, after outlier rejection, we used K-means clustering and then naive Bayes-based kernel density estimation

Figure 5

Results from using the K-nearest-neighbor algorithm to fill in missing and empty data

Figure 6

Outliers associated with low-density lipoprotein

Figure 7

(a) The Kernel density estimation performance is better than that of the normal distribution. For the AGE attribute and (b) for the TG attribute. In addition, we see that both attributes have different distributions in different regions. In addition, we used a clustering method to divide the data of a class into several clusters with the same statistical similarity in each cluster

Figure 8

Diabetes can be predicted using all of these features (13). (a). In the first scenario, that is, the presence of an outlier, M1 had a greater AAC in predicting diabetes than did the other models. (b). In the second scenario, after outlier rejection, M2 had a greater AAC for predicting diabetes than did the other models

Figure 9

Diabetes can be predicted using all features without blood glucose (12). (a). In the first scenario, that is, the presence of an outlier, M1 had a greater AAC in predicting diabetes than did the other models. (b). In the second scenario, after outlier rejection, M2 had a greater AAC in predicting diabetes than did the other models, and compared to the other experiments, in this case, the AAC decreased

Figure 10

Diabetes can be predicted by using principal component analysis to reduce dimensionality. Compared to other experiments, in this case, the highest AAC value is obtained. (a) In the first scenario, that is, the presence of an outlier, M1 had a greater AAC in predicting diabetes than did the other models. (b). In the second scenario, after outlier rejection, M2 had a greater AAC in predicting diabetes than did the other models, and compared to the other experiments, in this case, the AAC decreased

Figure 11

Changes in the number of first-class clusters can have an impact on the performance of the two models in different scenarios. The number of first- and second-class clusters has less impact than the number of first-class clusters

Figure 12

Comparison of diabetes prediction models in terms of AAC performance criteria