Diabetes is among the most prevalent chronic diseases worldwide, affecting millions of individuals and placing a substantial burden on patients, healthcare systems, and public health programs. It is characterized by elevated blood glucose levels resulting from insufficient insulin production, impaired insulin action, or a combination of both mechanisms. If inadequately controlled, diabetes can contribute to serious complications, including cardiovascular disease, kidney failure, neuropathy, vision loss, and other adverse outcomes. As the prevalence of diabetes continues to increase, greater attention is being directed toward the use of digital technologies, health data analytics, and artificial intelligence to support early detection, disease management, and clinical decision-making.

Diabetes informatics is an interdisciplinary field that applies health information systems, data analytics, computational methods, and digital technologies to diabetes prevention, diagnosis, treatment, and management. It involves the collection, storage, integration, analysis, and interpretation of diabetes-related information, including blood glucose measurements, demographic characteristics, clinical variables, medication data, comorbidities, and longitudinal health records. Analysis of these data can support personalized care, monitoring of disease progression, early identification of complications, predictive modeling, and evidence-based clinical decision-making.

De Melo and St. Rose [1] introduced the PM Generative AI framework for diabetes classification, demonstrating that generative artificial intelligence can improve predictive performance while maintaining model interpretability. Kadhm et al. [2] proposed a diabetes prediction system that integrates K-means clustering with a novel classification approach. Their methodology first grouped patients with similar clinical characteristics using unsupervised learning before applying classification algorithms, resulting in improved prediction accuracy compared with conventional classifiers. Sneha and Gangil [3] investigated optimal feature selection for early diabetes prediction. Zhu et al. [4] enhanced logistic regression by integrating principal component analysis (PCA) with K-means clustering. PCA reduced feature redundancy, while clustering identified hidden structures within the dataset before classification. Sinaga and Yang [5] provided a comprehensive review of the K-means clustering algorithm and proposed improvements that increase clustering stability and convergence. Wee et al. [6] compared numerous machine learning and deep learning techniques for diabetes detection. Their investigation included support vector machines (SVM), random forests, artificial neural networks, convolutional neural networks, and deep learning architectures.

Shakeel et al. [7] developed a cloud-based framework for diabetes diagnosis using K-means clustering. Ramadhan et al. [8] examined the effect of preprocessing techniques on Random Forest performance for type 2 diabetes detection. Saxena et al. [9] performed a comparative evaluation of multiple machine learning algorithms for diabetes detection, including logistic regression, decision trees, support vector machines, random forests, and ensemble methods. Wang and Su [10] proposed an improved K-means clustering algorithm that addresses sensitivity to initial cluster centers. De Melo and Davtyan [11] demonstrated the effectiveness of support vector machines for high-accuracy breast cancer classification. Although focused on oncology, the study illustrated the strong predictive capabilities of SVMs for clinical decision support and highlighted their applicability to other disease classification problems, including diabetes. Mostafa and Amano [12] presented an adjustable Round Robin scheduling algorithm based on clustering techniques. Although designed primarily for computer resource management, their work demonstrated how clustering methods can optimize computational efficiency, an important consideration for large-scale healthcare analytics and cloud-based medical applications. Sen et al. [13] presented a comprehensive survey of supervised classification algorithms, reviewing decision trees, naïve Bayes, support vector machines, neural networks, ensemble methods, and other classifiers. Cervantes et al. [14] conducted an extensive survey of support vector machine classification. Their review discussed kernel functions, optimization strategies, multiclass classification, and numerous biomedical applications. The authors concluded that SVM remains one of the most effective classifiers for high-dimensional medical datasets. Singh and Jaiswal [15] applied SVM combined with Whale Optimization Algorithm (WOA) within a Hadoop MapReduce framework. Although their application focused on audio signal classification, the study demonstrated that optimization algorithms can significantly improve SVM performance while supporting efficient big data processing. Luengo et al. [16] emphasized the critical role of data preprocessing in transforming raw big data into meaningful information. Their work discussed data cleaning, normalization, feature extraction, dimensionality reduction, missing-value treatment, and data integration, establishing preprocessing as an essential component of successful machine learning applications in healthcare.

Ingle et al. [17] demonstrated how big data analytics can identify patterns associated with vector-borne disease spread. Their research illustrated the effectiveness of machine learning and large-scale health data analysis for disease surveillance, providing methodological insights that can also benefit diabetes prediction research. Nilashi et al. [18] developed a soft computing approach for diabetes classification by combining fuzzy logic, clustering, and neural network methodologies. Their hybrid model effectively handled uncertainty and imprecise clinical information, producing improved classification accuracy compared with several conventional approaches. De Melo [19] presented a comprehensive overview of public health informatics, emphasizing the integration of health information systems, predictive analytics, artificial intelligence, and clinical decision support technologies. The book highlights the growing role of machine learning and data science in improving disease surveillance, healthcare management, and population health outcomes. Ahamad and Bharti [20] applied fuzzy logic to COVID-19 prevention strategies in India. Although focused on infectious disease, their study demonstrated the flexibility of fuzzy inference systems for modeling uncertain healthcare environments, supporting the broader application of fuzzy methodologies in medical decision support. Nguyen and Kreinovich [21] provided one of the foundational discussions of fuzzy logic applications in medicine. Their work explained how fuzzy systems accommodate uncertainty inherent in clinical decision-making and described numerous medical applications involving diagnosis, treatment planning, and patient monitoring. Finally, De Melo [22] developed a diabetes prediction model using electronic health records (EHRs). The study demonstrated that machine learning applied to routinely collected clinical data can accurately predict diabetes while supporting early intervention and clinical decision support. The research further emphasized the growing importance of EHR-based predictive analytics in modern healthcare and reinforced the value of integrating artificial intelligence into routine clinical practice.

Collectively, these studies demonstrate the evolution of diabetes prediction from traditional statistical methods to sophisticated artificial intelligence systems that integrate clustering, feature selection, ensemble learning, deep learning, fuzzy logic, and electronic health records. The literature consistently indicates that prediction performance depends not only on classifier selection but also on effective preprocessing, feature engineering, dimensionality reduction, and the integration of multiple analytical techniques. These advances provide a strong foundation for developing robust and clinically applicable diabetes prediction systems capable of supporting early diagnosis and personalized healthcare.

Diabetes informatics combines health data management, statistical analysis, artificial intelligence, and machine learning to support disease prediction, clinical decision-making, treatment planning, and population-level surveillance. Predictive models can analyze multiple patient characteristics simultaneously and identify complex relationships that may not be apparent through conventional statistical analysis alone.

Machine-learning methods have increasingly been applied to diabetes prediction and classification. Previous studies have investigated logistic regression, support vector machines, decision trees, random forests, K-means clustering, deep learning, and ensemble learning approaches. These studies demonstrate the potential of computational methods to support early diabetes detection. However, reported improvements in classification performance must be interpreted carefully because model performance can be influenced by data preprocessing, class imbalance, feature engineering, model selection, and experimental design.

Class imbalance represents an important challenge in diabetes classification. When non-diabetic observations outnumber diabetic observations, a machine-learning classifier may favor the majority class. Such a model may achieve acceptable overall accuracy while failing to identify a substantial proportion of patients with diabetes. In a clinical screening context, this problem is particularly important because false-negative classifications may delay further diagnostic evaluation and appropriate intervention.

The Synthetic Minority Oversampling Technique (SMOTE) is widely used to address class imbalance by generating synthetic minority-class observations from relationships among existing training samples. When correctly implemented, SMOTE can improve minority-class learning and increase sensitivity. However, oversampling must be carefully confined to the model-development process. If synthetic observations are generated before the train-test separation, or if augmentation affects the independent evaluation cohort, information leakage may occur and reported model performance may become overly optimistic.

A related methodological concern is the independence of the test population. Models cannot be meaningfully compared when they are evaluated using test sets with different sizes or class distributions. An apparent improvement in accuracy, sensitivity, or F1-score may result from changes in the evaluation population rather than genuine improvement in model generalization. Therefore, all competing methods should be evaluated using the same original, untouched test cohort.

XGBoost is a gradient-boosting approach that constructs an ensemble of decision trees sequentially, with each new tree attempting to correct errors made by the existing ensemble. Its ability to model nonlinear relationships and complex interactions among predictors makes it suitable for structured clinical data. Nevertheless, conventional XGBoost performance depends on the selection of hyperparameters, including the number of trees, maximum tree depth, learning rate, subsampling ratio, and feature-sampling ratio. Consequently, systematic hyperparameter optimization within the training data is necessary to develop a robust model while protecting the independence of the final test evaluation.

In this study, we introduce NSU-XGBoost, a leakage-controlled, SMOTE-enhanced, and hyperparameter-optimized XGBoost framework for diabetes detection. The framework integrates three methodological principles: strict isolation of an independent test cohort, minority-class augmentation confined to the training workflow, and cross-validated optimization of XGBoost hyperparameters.

Unlike experimental designs that modify or expand the test population through augmentation, NSU-XGBoost preserves the original test cohort throughout model development. Synthetic minority observations are introduced only within the training workflow, while model hyperparameters are selected using five-fold stratified cross-validation. The resulting optimized model is then evaluated on the frozen independent test cohort.

The primary objective of NSU-XGBoost is not simply to maximize overall accuracy. Rather, the framework seeks to improve the identification of diabetes-positive cases and reduce false-negative classifications while explicitly measuring the associated tradeoff in false-positive predictions and specificity. Accordingly, model performance is evaluated using confusion matrices, accuracy, sensitivity, specificity, precision, F1-score, ROC-AUC, and PR-AUC.

The objective of this study was to determine whether NSU-XGBoost could improve diabetes detection compared with tuned XGBoost without augmentation under an identical leakage-controlled experimental design. Both approaches were evaluated using the same frozen independent test cohort. This design allows observed differences in performance to be attributed to the modeling strategy rather than changes in the size or composition of the test population. By combining training-only class balancing, cross-validated model optimization, and rigorous independent evaluation, NSU-XGBoost provides a reproducible framework for developing diabetes classification models with greater emphasis on clinically important disease detection and transparent assessment of classification tradeoffs.

The dataset contained 768 observations with eight clinical predictor variables and a binary diabetes outcome. The predictors included pregnancies, glucose, blood pressure, skin thickness, insulin, body mass index, diabetes pedigree function, and age. The complete dataset was divided into two mutually exclusive groups:

The independent test cohort contained 100 non-diabetic and 54 diabetic observations. This test cohort was locked before model development and remained unchanged throughout feature engineering, hyperparameter selection, and threshold optimization.

Table 1 shows the first 10 rows of the diabetes data set.

Table 1. 10 first rows of the Indian PIMA diabetes data.

Table 2 depicts descriptive statistics of the data showing the total number of patients is 768.

Table 2. Descriptive statistics of the data.

Figure 1. The heatmap shows that glucose has the highest impact on the outcome.

A correlation heatmap (Figure 1) was generated using the complete dataset of 768 observations for descriptive and exploratory purposes. The heatmap was used only to characterize pairwise relationships among study variables and did not guide feature selection, preprocessing parameters, model tuning, or threshold optimization. All predictive modeling procedures were conducted separately under a leakage-controlled training and independent testing framework.

Figure 2 shows the distribution of glucose level categories by diabetes outcome. The figure compares the frequency of individuals with and without diabetes across four glucose level categories (Low, Normal, Pre-diabetes, and High). Non-diabetic participants (Outcome = 0) are predominantly represented in the Normal and Pre-diabetes categories, whereas diabetic participants (Outcome = 1) are concentrated in the High glucose category. The results demonstrate a strong positive association between elevated glucose levels and the likelihood of diabetes, highlighting blood glucose as one of the most important predictors for diabetes classification.

Figure 2. Histogram for the Glucose feature of the data set spread over 4 categories of diabetes.

Figure 3. Histogram for the BMI feature of the data set spread over 4 categories of diabetes.

Figure 3 shows the distribution of diabetes outcomes across BMI categories. The figure compares the number of individuals with negative (Outcome = 0) and positive (Outcome = 1) diabetes outcomes across underweight, normal-weight, overweight, and obese BMI categories. Positive diabetes outcomes are rare in the lower BMI categories and become increasingly frequent among overweight and obese individuals. The obese category contains the largest number of diabetes-positive cases, illustrating the association between elevated BMI and diabetes occurrence in the study population.

Figure 4 shows the distribution of glucose levels according to diabetes outcome. Kernel density distributions compare glucose levels between individuals without diabetes (Outcome = 0) and those with diabetes (Outcome = 1). The diabetes-positive group shows a clear rightward shift toward higher glucose values, whereas the diabetes-negative group is concentrated primarily around lower glucose levels. Although the distributions overlap, their distinct peaks and shapes demonstrate a strong association between elevated glucose levels and diabetes outcome.

Figure 4. The Glucose feature for 2 outcomes “diabetes” and “no diabetes”.

Extreme Gradient Boosting, commonly known as XGBoost, is a supervised machine-learning algorithm based on gradient-boosted decision trees. The method was selected for diabetes classification because clinical datasets frequently contain nonlinear relationships and complex interactions among predictors. Variables such as glucose, BMI, age, insulin, and family history may interact in ways that are not adequately represented by simple linear decision boundaries.

Unlike a single decision tree, XGBoost constructs an ensemble of trees sequentially. Each new tree is trained to improve the prediction errors of the existing ensemble. The final prediction is obtained by combining the contributions of all trees in the model:

where y i is the prediction for observation i, x i represents the predictor variables, K is the total number of trees, and f k represents the prediction generated by the k-th decision tree.

For binary diabetes classification, the model estimates the probability that an individual belongs to the positive diabetes class:

where F(X) represents the combined output of the boosted decision trees. The predicted probability is subsequently converted into a binary classification according to a selected decision threshold T:

The conventional threshold is T = 0.50. However, the default threshold was not assumed to be clinically optimal in this study. Threshold optimization was therefore treated as a separate stage of model development and was performed using training data only.

This study developed NSU-XGBoost, a leakage-controlled framework that integrates training-only SMOTE augmentation, cross-validated XGBoost hyperparameter optimization, and adjustable probability thresholds for diabetes classification. By maintaining a frozen independent test cohort throughout model development, the framework avoids information leakage and permits a more reliable assessment of model performance.

The results demonstrate that NSU-XGBoost can substantially improve the detection of diabetes-positive cases. SMOTE-based balancing of the training data enables the model to better learn patterns associated with the minority diabetes class, while adjustment of the probability threshold provides additional control over the sensitivity-specificity balance. Lowering the threshold increases sensitivity and reduces false-negative classifications, whereas increasing the threshold improves specificity and reduces false-positive classifications. Thus, the optimal operating threshold depends on the intended clinical application rather than on overall accuracy alone.

For symptomatic or high-risk patients, a sensitivity-oriented threshold may be appropriate because the primary objective is to minimize missed diabetes cases and identify patients requiring further clinical evaluation. However, the threshold should not be selected arbitrarily or adjusted after examining the independent test results. Instead, a predefined threshold should be selected using training-only cross-validation to achieve a clinically appropriate sensitivity target while retaining the highest possible specificity. For example, a screening application could select the training-derived threshold that achieves a prespecified sensitivity of 90% or 95%.

Conversely, in applications where unnecessary follow-up testing and false-positive classifications are of greater concern, a higher threshold may be selected to favor specificity. This flexibility allows NSU-XGBoost to operate at different clinically defined decision points. The framework therefore does not provide a single universal threshold for all patients and settings, but rather a systematic method for selecting an operating point appropriate to the intended use.

Importantly, NSU-XGBoost should be considered a clinical decision-support and risk-classification framework rather than a replacement for established diagnostic procedures. Patients identified as high risk, particularly those presenting with symptoms consistent with diabetes, should undergo appropriate clinical assessment and confirmatory laboratory testing.

The selected threshold 0.1 was chosen to illustrate the clinical tradeoff between sensitivity and specificity rather than to identify a globally optimal operating point. We have revised the manuscript to acknowledge this limitation and now state that future work will include threshold scanning using cross-validation and predefined clinical criteria such as the Youden index or target sensitivity.

The model should only flag risk. Diagnosis should rely on established laboratory criteria. Under the 2026 ADA Standards, classic hyperglycemic symptoms together with a random plasma glucose of at least 200 mg/dL can establish diabetes; other diagnostic pathways include A1C, fasting plasma glucose, or oral glucose-tolerance testing.

Overall, the findings demonstrate that the value of NSU-XGBoost lies not only in classification performance but also in its ability to provide a controlled and transparent balance between sensitivity and specificity. By combining leakage-controlled training, minority-class augmentation, optimized XGBoost modeling, and clinically oriented threshold selection, NSU-XGBoost offers a flexible framework for diabetes screening and risk stratification. Future research should validate the framework in larger and external patient populations and investigate whether population-specific or clinical-context-specific operating thresholds can further improve its utility in real-world healthcare settings.