An intelligent diabetes classification and perception framework based on ensemble and deep learning method

Diabetes prediction on PIMA data set

A predictive model based on machine learning and deep learning utilizes risk factors as input features to predict whether patients have diabetes. This computer-aided system assists physicians in decision-making. Haritha, Babu & Sammulal (2018) employed a hybrid approach, predicting diabetes patients using Firefly and cuckoo search techniques for feature selection. Among these techniques, Cuckoo-Fuzzy-KNN achieved the highest accuracy. Maniruzzaman et al. (2018) compared six feature selection techniques (principal component analysis, analysis of variance, Fisher discriminant ratio, random forest, logistic regression, mutual information) and ten machine learning classifiers on a dataset of 768 diabetes patients. Benbelkacem & Atmani (2019) used an RF classifier to predict diabetes patients and compared its performance with other classifiers, finding a lower error rate than SVM and C4.5. Reddy et al. (2020) employed NB, RF, SVM, GB, KNN, and logistic regression (LR) to classify diabetes in PIMA Indian patients.

To extract the most useful features from the diabetes data set, the research authors (Arora et al., 2022) employed the K-means clustering technique. To classify people with diabetes, they used an SVM Classifier on the collected features. According to the findings of this study, this method has a 98.7% accuracy.

Using the patient’s QR card, researchers in the study (Marzouk, Alluhaidan & El Rahman, 2022) developed a system for monitoring diabetes and communicating treatment updates to medical professionals. The author performed an experiment on PIMA and diabetes synthetic data sets. RF, NB, LR, and ANN are all utilized in the diabetes categorization. The experimental result shows that the ANN achieved 81.6% accuracy. In a article (Patil & Ingle, 2021), KNN, NB, SVM, and RF classifiers for the classification of diabetes are compared, and the results are compared with ANN. Their proposed method achieved 97.66% accuracy for the classification of diabetes. A study (Jaganathan et al., 2022) used KNN, RF, NB, and LR to classify diabetes. RF achieved 89.58% accuracy, which is the highest compared to another method. In a study, the author (Chang et al., 2023) used a DT, NB, and an RF classifier to classify diabetes. The prediction accuracy of random forest is 85.17%.

Diabetes prediction with neural network

To acquire precise prediction results for a real-world problem, neural networks can be used in many contexts. Its versatility means it may be used in various contexts where it will perform admirably and yield the desired results for solving practical problems. For instance, Pradhan et al. (2020) used a Neural Network to classify the patients as sick or healthy 768 diabetes patients. The study employed a standard backpropagation method for ANN model training. The experimental results verified that ANN classifies patients with an 87% accuracy rate. Optimal feature selection plays a vital role in predictive modeling. Kannadasan, Edla & Kuppili (2019) selected optimal features using a stacked autoencoder from the PIMA 768 data set and applied DNN for prediction. The model is evaluated using precision, recall, specificity, and F1 measures. The experimental results showed that this model achieved 86.26% accuracy, 87.92% recall, and 90.66% precision. Naz & Ahuja (2020) applied deep neural networks (DNN), NB, and DT. The experimental results show that deep learning achieved the highest accuracy compared to another classifier used in this proposed work with 98% accuracy. Ayon & Islam (2019) also applied a DNN from the Indian PIMA data set to detect diabetes. The model’s performance is evaluated using two measurement parameters, F1 and Accuracy.

The blood vessels in the retina can be damaged due to diabetic retinopathy. Alfian et al. (2020) utilized a DNN to predict diabetic retinopathy in patients, employing recursive feature selection and elimination techniques to eliminate irrelevant attributes. The experimental results demonstrated that the DNN predicted retinopathy with an accuracy of 82.03%. Similarly, Gadekallu et al. (2020) predicted retinopathy from risk factors using DNN, extracting significant features with principal component analysis and normalizing raw data with a standard scalar. The authors also reduced the dimension of data using the Firefly algorithm. Early detection of end-stage renal disease can benefit patients with kidney disease, protecting them from developing adverse health conditions. In clinical trials, Belur Nagaraj et al. (2020) utilized LR, SVM, RF, feed-forward neural network (FFNN), Cox proportional hazard regression, and the KFRE model. The proposed FFNN approach outperformed the baseline scheme. Kowsher et al. (2019) compared ANN, RF, NB, and DT classifiers using information from 9,483 diabetic patients. In another study (Tan et al., 2022), the authors proposed a genetic algorithm (GA) based on a stacking technique for diabetes prediction. They selected relevant features using a GA, applied neural network and SVM as baseline learners for meta-models, and conducted experiments on the Early-Stage Diabetes Risk Prediction dataset. Their proposed method achieved a remarkable 98.71% accuracy on the Early-Stage Diabetes Risk Prediction dataset.

Ensemble learning and decision tree for diabetes prediction

Abedini, Bijari & Banirostam (2020) applied an ensemble hierarchical model for diabetic patient classification. Firstly, LR and DT are trained, and the output is fed to a neural network. This approach classifies the diabetes patients with 82% accuracy from the PIMA diabetes data set. Pei et al. (2019) identify the potential of Type II diabetes using a DT classifier to identify diabetes. The experimental results show that the proposed model achieved 94.8% AUC, 94.2% accuracy, 94.0% precision, and 94.2% recall value. Maniruzzaman et al. (2020) identified the risk factors for diabetes disease based on p-value and odd ratios. Adaboost, NB, RF, and DT were applied to predict the diabetes patients. The experimental results show that LR and RF combination perform better.

Sneha & Gangil (2019) used optimal features for early diabetes detection and applied RF, SVM, NB, DT, and KNN on 2,500 data items. The experimental results show that SVM achieved the highest accuracy, which is 77.73%. Nibareke & Laassiri (2020) compared DT, LR, and NB classifiers on Vincent Sigillitan’s data set. This work aims to overview machine learning modeling and big data tools. The experimental results show that the decision tree achieved 0.766% accuracy and 0.125 RMS value.

Datasets

This study used two datasets for diabetes prediction.

Data preprocessing

Data preprocessing is the procedure of cleaning and transforming raw data. This work uses label encoding to convert categorical labels into discrete forms. Additionally, we normalize the PIMA dataset using the standard scaler normalization technique (Sharma et al., 2021). The equation for the standard scaler is provided below:

(1) $$Z=Xi-\mu /\sigma$$ $\mu$ is a mean of a given data and $\sigma$ is variance of a given data. While X_i is an input feature.

Feature selection strategy on Pima dataset

In the PIMA dataset, we have eight features, and we select the best six features using the SFS backward strategy. Figure 2 details the selected and dropped features.

Figure 2 displays details of all features and the selected features from the PIMA dataset. Selected attributes are highlighted with their respective colors.

Feature selection strategy on early risk diabetes dataset

In the Early Risk Diabetes dataset, we have 16 features, and we selected the best 12 features using the SFS backward strategy. Figure 3 illustrates the details of the selected and dropped features.

Figure 3 displays the selected features from the Early Diabetes Risk dataset, with each attribute highlighted in its respective color.

Machine learning method for diabetes prediction

ANN, RF, Tab-net, SVM, and GB are employed for diabetes prediction, and the parameters of each method are tuned using grid search. Grid search is a hyperparameter tuning method that optimizes the hyperparameters of each method. In grid search, we first define a range of hyperparameters, and the machine learning model is evaluated for possible combinations. The hyperparameter yielding the best performance on the validation data is then selected for the final training of the model.

Random forest

RF is an ensemble learning classifier consisting of multiple decision trees, proposed by Breiman and Adele Cutler (Khan et al., 2022). RF is widely used in various application areas, including medical diagnosis (Solaiyappan & Wen, 2022). In this work, we employed RF for diabetes prediction. Figure 4 illustrates the workflow steps of random forest.

In Step 1, we divide the dataset into training and testing sets. The bootstrapping method creates new datasets from the training set. In Step 2, a decision tree is constructed based on the results of Step 1. In Step 3, multiple DT are built to form a random forest, and Steps 1 and 2 are repeated. In this step, each tree in the forest votes for the given variables, and the mean vote for each class is computed. RF has various hyperparameters requiring optimal tuning of results. We train RF for diabetes prediction using these parameters, as discussed in Table 3. The rationale for choosing this method is based on the assumption that Random Forest performs well with tabular data, such as the PIMA and early-risk diabetes datasets.

Table 3:

Hyperparameter of random forest for diabetes prediction.

Hyper parameter	Values
Random state	None
Min samples leaf	1
Min samples split	2
Max features	Auto
Bootstrap	True
N estimators	50
Criterion	Gini

DOI: 10.7717/peerj-cs.1914/table-3

Evaluation metrics for diabetes prediction

When training classifiers to address practical problems, they learn from the given environment and make predictions. Classifier performance evaluation is essential to determine the accuracy of class label predictions. Various evaluation metrics are utilized in diabetes prediction, including accuracy, precision, recall, and F-measure.

Correlation analysis

Figure 5 displays the correlation heatmap of the PIMA diabetes dataset. It indicates that the blood pressure feature has a low correlation with the outcome variable, while another feature exhibits a high correlation with the target class label.

Figure 6 displays the correlation heatmap of features in the early-stage diabetes risk dataset. The gender feature shows a low correlation with the target attribute, while another feature exhibits a high correlation with the target class label.

Experimental results without feature selection

Table 6 presents the results of machine learning techniques for diabetes prediction on the PIMA and early diabetes risk datasets using all features. The ANN achieved an accuracy of 98.71%, and the RF achieved 96.79% accuracy on both the PIMA and Early-Risk Diabetes datasets with all available features. It suggests that ANN and RF outperform other models in predicting both datasets. However, the TabNet model exhibits lower performance on the early-risk diabetes dataset than other methods while showing better prediction results than RF, GB, and SVM on the PIMA dataset. These variations in model performance are attributed to their learning capacities, with ANN and RF demonstrating better capabilities in learning complex decision boundaries for intricate non-linear problems.

Figure 7 displays a comparative graph of results for machine learning methods on the PIMA and Early Risk diabetes datasets without feature selection. The graph contrasts the accuracy and F score of different machine learning methods, revealing that the ANN outperformed other methods on the PIMA dataset. At the same time, RF excelled on the Early Risk Diabetes dataset.

The confusion matrices of the machine learning model for both datasets are presented in Fig. 8 without feature selection. It indicates that RF has a low misclassification rate on the Early Risk Diabetes dataset. However, the Tab-net model exhibits a lower misclassification rate on the PIMA dataset than other machine learning models.

Experimental results of chi-square feature selection

The experimental results of the chi-square feature selection are presented in Table 7. It illustrates that the prediction accuracy of ANN and RF is 98.69% and 98.07%, respectively, on the PIMA and Early Diabetes Risk datasets. The performance of both models has improved compared to the performance without feature selection. On the other hand, the performance of some models has decreased on the selected features of the chi-square feature selection. The reason for the reduction in the performance of some methods is that the learning capacity of ML models varies due to input attributes and their relationship with the target class label. The overall prediction performance of ANN and RF with the chi-square feature selection is noteworthy.

Figure 9 demonstrates the machine learning results using the chi-square feature selection. It indicates that the prediction performance of ANN and RF is high on both datasets, respectively. With this feature selection method, most machine learning models show improved prediction results on both datasets.

Experimental results of MI feature selection

The experimental results of MI feature selection are described in Table 8. It demonstrates that the machine learning model’s performance is improved on both datasets with this feature selection method compared to the chi-square feature selection. The prediction accuracy of ANN and RF is also enhanced in both datasets.

The experimental results of machine learning models with MI feature selection are presented in Fig. 10. This figure illustrates that the machine learning model’s performance is improved in terms of accuracy and F-score with the MI feature selection method compared to chi-square and without feature selection.

Experimental results of sequential feature selection

Machine learning methods are applied to the most relevant features, and the performance of each model is evaluated. Table 9 presents the prediction results for both datasets using RF, GB, ANN, TabNet, and SVM. The ANN classifier detects diabetes patients with 99.35% accuracy on the PIMA dataset, and the RF classifier identifies early diabetes patients with 99.36% accuracy, as shown in Table 9. The prediction performance of all other models is improved with this feature selection method on both datasets. The intuition behind this improvement is that SFS identified more informative features, enhancing the classification results of machine learning models.

Figure 11 illustrates the comparison graphs of machine learning methods based on accuracy and F-score measures for the PIMA and Early Risk Diabetes datasets. It indicates that the features selected by SFS exhibit better prediction performance for all methods than all feature sets. It suggests that selecting relevant features enhances prediction performance. This intuition arises from the fact that the selected features have a strong relationship with the target class label, and reducing the number of features helps diminish noise in the data. The results highlight that ANN and RF outperform other methods on the PIMA and Early Risk Diabetes datasets.

The confusion matrices of the machine learning models for both datasets are presented in Fig. 12 with the sequential feature selection. It demonstrates that the machine learning models have a lower misclassification error on both datasets with the sequential feature selection than without feature selection. The confusion matrices of machine learning models with the chi-square and mutual information feature selection are presented in Figs. 13 and 14, respectively.

Comparison of the proposed method with the existing methods