Diabetes mellitus is a severe and chronic disease characterised by metabolic disorders in which the pancreas either fails to produce insulin, or the body cannot effectively utilise the insulin produced (Sneha and Gangil, 2019). Lack of awareness about the symptoms and complications of diabetes is prevalent due to limited healthcare resources in many parts of the world (Webber, 2013). There are approximately 40 different types of diabetes, with some common types being Type 1 (insulin-dependent), Type 2 (insulin-independent), gestational diabetes, and pre-diabetes (Kharroubi and Darwish, 2015).

According to statistical reports from various healthcare organisations, it is estimated that globally, 463 million adults, which accounts for 9.3% of the population aged between 20 and 79 years, are affected by this chronic disease (Diabetes Federation International and IDF, 2019). This highlights the widespread prevalence and significance of diabetes as a global health issue.

Projections suggest that the prevalence of diabetes will continue to increase significantly, with an estimated 578 million individuals affected by 2030. According to the Diabetes Atlas 2019 by the International Diabetes Federation (IDF), approximately 50% or 231 million people living with diabetes remain undiagnosed and unaware of their condition due to limited healthcare resources (Diabetes Federation International and IDF, 2019).

In 2019 alone, diabetes was responsible for 4.2 million deaths worldwide. This chronic disease can have detrimental effects on various organs in the human body, including the brain, nerves, heart, kidneys, eyes, and skin. Recognising the symptoms and signs of diabetes is crucial for early detection and management. Some common early symptoms observed in individuals with diabetes or those at risk include excessive thirst, fatigue, unexplained weight gain, dizziness, skin discoloration, sexual dysfunction, fungal infections, high blood sugar levels, and frequent urination (Sneha and Gangil, 2019). These symptoms serve as important indicators for seeking medical attention and further evaluation.

Indeed, given the significant impact and global burden of diabetes, there is an urgent need to leverage computational intelligence techniques for improved prediction and prevention of this disease. By utilising advanced machine learning and artificial intelligence algorithms, we can develop models that can effectively identify individuals at risk of developing diabetes. These models can analyse large-scale datasets, extract meaningful patterns, and generate accurate predictions.

The application of computational intelligence techniques in diabetes prediction can have several benefits. Firstly, it can enable early disease detection, allowing for timely intervention and management strategies. This early detection can aid in preventing or delaying diabetes-related problems, improving overall health outcomes for people.

Furthermore, by accurately predicting diabetes, healthcare professionals can implement preventive measures and provide personalised care plans for high-risk individuals. This can involve lifestyle modifications, dietary interventions, exercise regimens, and medication management to effectively manage and control blood sugar levels.

Overall, applying computational intelligence techniques to diabetes prediction can significantly enhance medical results, lessen the condition’s toll, and encourage proactive and preventative healthcare practices for those at risk.

However, healthcare data are growing drastically, and the traditional machine learning approaches have been found inadequate to handle such voluminous data for accurate disease predictions. Ensemble learning techniques offer better performance in this regard.

This work aims to create a model that accurately predicts diabetes using ensemble learning approaches. Our work’s contribution is as follows:

• Performing exploratory data analysis to improve the dataset’s quality assessment.

The rest of the paper is organised as follows. Related work is discussed in Section 2. The adopted methodology and the dataset are presented in Section 3. Then, the experimental details and results are described and analysed in Section 4. Next, the comparative analysis with existing similar works is presented in Section 5. Lastly, the conclusion and the future direction of the research are provided in Section 6.

In recent years, copious work has been done on the prediction of diabetes using machine learning and ensemble learning tools and techniques (Ganie et al., 2022a; Ganie and Malik, 2022a). Different datasets, algorithms, and methodologies used by the researchers to carry out this research work have been discussed. The developed models have yielded better results and can be used to support healthcare providers in data-driven decision-making. This section reviews some of the key relevant papers on applying ensemble learning approaches to forecast diabetes.

Li et al. (2020) developed a model to predict diabetes using ensemble learning techniques to enhance disease prediction using the Pima diabetes dataset. They achieved the highest results with extreme gradient boosting (XGBoost), with an accuracy rate of 80.20%. The authors proposed the improved feature combination classifier using the XGBoost model, which can be explored to better predict diseases in the healthcare industry. Mahabub (2019) tested different ensemble learning techniques, such as AdaBoost, gradient boost, XGBoost, random forest, etc., to predict diabetes, considering several clinical parameters such as pregnancy, skin thickness, glucose, insulin, blood pressure, diabetes pedigree function, body mass index (BMI), age, and class variable (outcome). They achieved the highest accuracy rate of 84.42% with the multilayer perceptron algorithm. Mushtaq et al. (2022) proposed an optimised model using a voting classification based on the ensemble method to predict diabetes using the Pima diabetes dataset. This research work used a two-stage model selection process to develop the model. The voting classifier reached the best accuracy rate of 81.50% among all the classifiers. Furthermore, Tomek and synthetic minority oversampling technique (SMOTE) techniques were used for data balancing to remove the biases from the dataset. The authors suggested that the research be continued to estimate the likelihood that nondiabetic patients will develop this condition in the future.

Beschi Raja et al. (2019) employed different boosting algorithms for the diabetes prediction model development. The gradient boosting algorithm attained the highest accuracy rate of 89.70% among all the classifiers. Other statistical measurements have also been evaluated to validate the proposed model. Khan et al. (2021) developed a model for diabetes prediction using boosting method. The authors explored different classifiers such as gradient boosting, hybrid k-nearest neighbour (kNN), j48, deep learning, naive Bayes, and artificial neural network (ANN) for predictive analytics. Among all the classifiers, the gradient boosting algorithm attained the best results. In addition, the results were validated using the k-fold cross-validation method. The authors suggested that this model can be used as a prognosis tool in the healthcare industry for early disease prediction. Lai et al. (2019) developed a complete framework for the predictive analysis of diabetes. The gradient boosting machine techniques were used with hyperparameter tuning, particularly for class balancing, which minimised the loss of prediction probabilities regarding classification.

Singh et al. (2021) introduced an ensemble approach based framework called eDiaPredict to forecast the diabetes status of patients. The proposed methodology incorporates XGBoost, random forest, support vector machine (SVM), neural network, and decision tree. The efficacy of eDiaPredict is demonstrated through its implementation on the PIMA Indian diabetes dataset, resulting in an attained accuracy, precision, and sensitivity of 95%, 88%, and 90.32%, respectively, with the combination of XGBoost and random forest. Hasan et al. (2020) presented a framework for predicting diabetes using kNN, decision trees, random forest, AdaBoost, Naive Bayes, XGBoost, and multilayer perceptron. They employed a weighted ensemble of the machine learning models to improve the prediction accuracy, and experimented on the PIMA Indian diabetes dataset. The proposed ensemble model achieved a significantly higher AUC and specificity of 0.950 and 0.934, respectively. However, it exhibited lower accuracy, precision and sensitivity of 88.84%, 84.32%, and 78%, respectively.

Figure 1 illustrates the procedural flow of the proposed framework employed in this experimental study. It outlines the sequential steps undertaken to enhance the prediction accuracy of diabetes using an ensemble learning technique based on boosting methods. The Pima Indians diabetes dataset, obtained from the Kaggle community, was utilised for this study. Initially, the necessary Python library packages were installed in Jupyter Notebook. Exploratory data analysis was conducted to enhance the dataset’s quality assessment. During this phase, missing values were identified and replaced through data imputation. The Interquartile Range method was applied to detect outliers in the dataset (Ganie et al., 2023).

Other necessary libraries were run to check the dataset for any corrupted data. Upsampling and normalising were also carried out before the five boosting methods under consideration were created. The dataset was split with a ratio of 80:20, where 80% of the data was employed for training the boosting algorithms, and 20% was used to test and validate their efficacy. Hyperparameter tuning was applied during the model-building process for better results.

The experimental minutiae and findings generated by the application of boosting algorithms for diabetes prediction are presented and discussed in this part. The outcomes obtained after using the suggested framework are methodically presented and examined. The evaluation is carried out thoroughly, considering several measures for the evaluated boosting algorithms, including accuracy, recall, precision, F1-score, micro-weighted score, average weighted score, and the receiver operating characteristic (ROC) curve. These measures give us important information about how well the boosting algorithms perform and how well they forecast diabetes.

Figure 13 presents a comparative analysis of the five boosting algorithms considered in the experiment. The algorithms were compared in terms of accuracy, AUC value, and runtimes. Among these algorithms, gradient boosting achieved the highest accuracy rate, reaching a maximum accuracy of 96.75%. Following gradient boosting, LightGBM achieved an accuracy rate of 94.15%, AdaBoost achieved 91.55%, CatBoost achieved 92.2%, and XGBoost achieved 93.5%. In addition, gradient boosting also excels in terms of AUC. However, in terms of runtime XGBoost outperforms others, requiring the least runtime.

We compared the highest accuracy achieved by our proposed method (i.e., using gradient boosting) with several relevant literature in terms of accuracy, as shown in Table 4. The implemented processes, such as data imputation for handling missing values, detection, and box plotting for outlier elimination, could be credited with the reason for achieving improved accuracy.

In this research, we investigated the effectiveness of five boosting algorithms, namely, XGBoost, CatBoost, LightGBM, AdaBoost, and gradient boosting, for predicting diabetes disease. Various preprocessing techniques, such as imputation, Z-score, and cleaning methods, were applied to improve the quality of the dataset. Additionally, to enhance disease prediction, data normalisation, upsampling, and hyperparameter tuning were performed.

According to the experimental findings, gradient boosting had the greatest accuracy rate of 96%. Additionally, it did well in terms of other evaluation criteria like ROC curve, precision, recall, and f1-score. The feature importance technique revealed how independent features contributed to the outcome of the final prediction.

Furthermore, when compared to similar related efforts, the suggested framework performed better than existing systems. Other ensemble learning strategies, such as bagging and stacking, can be added to further increase the quality of the outcomes. To increase the scope of this research, the proposed method can also be used for other healthcare datasets with comparable features.

In future studies, exploring deep learning techniques could lead to better detection and prediction of diabetes. These advancements in machine learning and deep learning can contribute to more accurate and efficient healthcare solutions.