Feature filtering

In this study, it is necessary to filter the dataset to eliminate features that are not relevant to the objectives. The features 'Location', 'StudyProject', and 'RecordID' are removed, as they do not provide meaningful information for the analysis of pediatric diabetes. 'Location' and 'StudyProject' are administrative variables without clinical, biochemical, or genetic relevance, while 'RecordID' serves merely as a unique identifier with no analytical value. Removing these features reduces dimensionality, thereby simplifying the dataset and enhancing both the efficiency and performance of the models. In the context of machine learning and data science, reducing dimensionality is a critical step that improves model interpretability and mitigates the risk of overfitting by eliminating noise and irrelevant data^[23][24].

Data incompleteness

Addressing missing values is essential in preparing medical datasets, where incomplete data can significantly impact the performance and reliability of machine learning models^[25]. In this study, data incompleteness is managed through mean imputation, a technique widely employed for handling missing data, particularly when dealing with continuous variables. This approach involves replacing missing values with the mean of the corresponding feature, thereby preserving the overall central tendency and distribution of the data^[26].

Mean imputation is applied by calculating the mean of each feature with missing values and substituting this mean for the missing entries^[27]. Mathematically, for a given feature Xj with missing values, the imputed value x^ij for any missing observation xij is given by Equation (1).

where n represents the total number of observations for which the feature Xj is not missing. This method assumes that the missing data are Missing At Random (MAR), implying that the probability of missingness is not related to the missing values themselves but may be related to the observed data. Mean imputation preserves the mean of the observed data and minimizes the impact on the overall distributional properties, maintaining the variance structure to some extent. However, it may reduce the natural variability of the dataset by introducing repeated values, which is an acceptable trade-off in this context due to the small proportion of missing data^[28][29][30].

While other imputation methods, such as median or mode imputation, can handle non-normally distributed data or categorical variables, mean imputation is chosen for its computational efficiency and its ability to maintain the central tendency of the dataset. This ensures that the imputations do not introduce significant bias, allowing the synthetic data generation models to learn from a complete and consistent dataset while preserving its original characteristics as closely as possible.

Data normalization

Data normalization is a fundamental step in preparing datasets for machine learning models, particularly neural networks, which are highly sensitive to the scale of input data^[31]. In this study, Min-Max Scaling is employed to normalize the continuous variables, transforming each feature to a fixed range, typically [0, 1]. This normalization technique is defined by Equation (2).

where X represents the original feature values, Xmin is the minimum value, and Xmax is the maximum value of the feature^[32]. Min-Max Scaling is chosen because it preserves the original distribution shape while ensuring that all features contribute equally to the model's learning process. This method is particularly advantageous over other normalization techniques, such as Z-score normalization, which centers the data around the mean and scales based on standard deviation but can be more sensitive to outliers and does not always maintain the relative spacing of values^[33][34].

The use of Min-Max Scaling is especially beneficial in the context of Generative Adversarial Networks (GANs) and Wasserstein GANs (WGANs). These models require a stable training process to effectively generate high-quality synthetic data, and normalization helps by ensuring that gradients do not vanish or explode due to disparate feature scales. By standardizing the input data, the models can learn underlying patterns more effectively, thereby improving the robustness and quality of the synthetic data generated^[35].

Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs), introduced by Goodfellow et al.^[36], are a powerful class of generative models widely used for creating synthetic data that closely mimics real-world data distributions. GANs consist of two neural networks: a generator G and a discriminator D, which are trained simultaneously through an adversarial process. As depicted in Figure 2, the generator's role is to produce synthetic data samples by transforming random noise, typically drawn from a prior distribution such as a Gaussian, into data that resembles the real dataset. The discriminator, on the other hand, acts as a binary classifier that attempts to distinguish between genuine data samples and those generated by the generator.

Figure 2 Overview of the Generative Adversarial Network (GAN) Architecture. 

The adversarial training process of GANs is formulated as a minimax game, where the generator aims to "fool" the discriminator by producing data that the discriminator classifies as real, while the discriminator strives to correctly identify whether a given sample is real or fake. The objective function that governs this process is defined by Equation (3).

where:

In this framework, the generator G learns to minimize the probability of the discriminator correctly identifying its synthetic outputs as fake, while the discriminator D aims to maximize the accuracy of its predictions. This continuous feedback loop drives both networks to improve iteratively: the generator becomes better at creating realistic data, and the discriminator becomes more adept at detecting synthetic samples. The training reaches equilibrium when the discriminator cannot distinguish between real and generated data, outputting a probability close to 0.5 for both^[36].

GANs rely on two critical components to function effectively:

Model design and implementation

To effectively generate synthetic datasets that capture the complex characteristics of pediatric diabetes data for this study, a GAN was carefully designed and implemented with specific architectural choices and training parameters. These were selected to achieve an optimal balance between model complexity, training stability, and the generation of high-quality synthetic data. Table 2 provides a detailed overview of the key architectural components and training parameters used in this study.

The architecture and training parameters of the GAN model were selected based on established practices in the literature and tailored to meet the requirements of this study. The generator and discriminator were designed with dense layers and activation functions, such as Tanh and Leaky ReLU, which are widely recognized for their ability to improve gradient flow and stabilize adversarial training^[36][38]. The RMSprop optimizer, combined with a carefully chosen learning rate, was employed to ensure stable convergence during training^[37]. To enhance model robustness and diversity in the generated data, additional techniques such as batch normalization, noise addition, label smoothing, and dropout were integrated. These approaches, supported by prior research, have proven effective in improving training stability and preventing overfitting in GAN architectures^[39][40]. Post-processing steps, including descaling and adjustment of discrete variables, were applied to ensure the synthetic datasets reflected realistic and clinically meaningful values. Collectively, these design choices contributed to the successful generation of synthetic datasets that accurately capture the complex characteristics of pediatric diabetes data.

Wasserstein Generative Adversarial Networks (WGANs)

Wasserstein Generative Adversarial Networks (WGANs) are an improved variant of traditional GANs, designed to address common challenges such as instability during training and mode collapse, where the generator fails to capture the full diversity of the real data distribution. Introduced by Arjovsky et al.^[35], WGANs use the Wasserstein distance (also known as Earth Mover's Distance) as a loss function, which provides more stable gradients for training and leads to improved convergence properties. The Wasserstein distance is defined as shown in Equation (6).

where:

Using this distance allows WGANs to maintain meaningful gradients throughout the training process, even when there is limited overlap between the real and generated data distributions. Instead of the discriminator outputting a probability, as in traditional GANs, the critic in a WGAN outputs a real-valued score, which estimates the Wasserstein distance between the real and generated data distributions. The generator is trained to minimize this distance, while the critic is trained to maximize it. This formulation leads to more stable training, as it avoids problems such as vanishing gradients that can occur with other types of loss functions^[41].

To ensure that the Wasserstein distance is valid, the critic function must satisfy the Lipschitz continuity condition, which is enforced by clipping the weights of the critic to a small range (e.g., [−0.01,0.01]). This weight clipping helps maintain stable gradients and promotes effective training, further reducing the likelihood of mode collapse and enhancing the quality of the synthetic data produced.

While WGANs introduce significant improvements in training dynamics through these changes, their overall architecture in terms of the core components (generator and critic) remains similar to that of a traditional GAN. Both models consist of a generator network that creates synthetic data and a discriminator (or critic in the case of WGAN) that evaluates the data. Therefore, the differences between GAN and WGAN primarily lie in the loss function and training strategies, rather than in the architectural layout itself^[41][42]. For this reason, a separate figure illustrating the architecture of WGAN is not included, as it would visually mirror the figure already provided for the standard GAN.

Model design and implementation

To effectively generate synthetic datasets that capture the complex characteristics of pediatric diabetes data for this study, a WGAN was carefully designed and implemented with specific architectural choices and training parameters. The chosen architecture and parameters aim to leverage the advantages of the Wasserstein distance, ensuring stable training and high-quality data generation. The details are shown in Table 3.

The choice of these parameters and configurations reflects a balance between model complexity, training stability, and the ability to generate high-quality synthetic data. The generator and critic architectures, with three dense layers and Leaky ReLU activations, are specifically designed to capture the intricate patterns in pediatric diabetes data. Batch normalization is incorporated in the generator to stabilize gradient updates, while dropout is applied in the critic to prevent overfitting during adversarial training^[37][41].

The RMSprop optimizer with a learning rate of 0.00005 was selected to ensure gradual convergence and prevent instability during the adversarial training process. This is particularly important in the WGAN framework, as the Wasserstein distance requires precise optimization to maintain the Lipschitz continuity condition^[35]. Weight clipping with a small range (0.01) enforces this condition, further enhancing the stability and effectiveness of the training process^[14].

Additionally, the number of training epochs (100) and the batch size (32) were carefully chosen to balance sufficient training iterations and computational efficiency. These configurations allow both the generator and critic to achieve optimal performance without overfitting the data. Post-processing adjustments, including descaling and the refinement of discrete variable values, ensure that the generated synthetic data remains realistic and clinically meaningful^[38][39].

The WGAN framework was chosen for its ability to address limitations in traditional GANs, particularly in stabilizing the training process and generating diverse, high-quality data. By combining the standard GAN approach with the more robust WGAN framework, this study leverages the strengths of each model to produce synthetic datasets that are not only accurate but also diverse and suitable for developing reliable predictive models.

Jensen-Shannon Divergence (JSD)

The Jensen-Shannon Divergence (JSD) is a statistical measure used to quantify the similarity between two probability distributions^[43]. As a symmetrized and smoothed version of the Kullback-Leibler divergence, the JSD provides a bounded and interpretable metric for assessing how one distribution diverges from another. It is particularly suitable for evaluating the performance of generative models, such as GANs and WGANs, by measuring the extent to which the synthetic data replicates the distributional characteristics of real-world data^[44][45].

Mathematically, the JSD between two discrete probability distributions P and Q is defined as shown in Equation (7):

where:

The JSD is symmetric (JSD(P || Q) = JSD(Q || P)) and is bounded between 0 and 1. A value of 0 indicates that the two distributions are identical, while values closer to 1 indicate greater divergence. This makes JSD a robust metric for comparing distributions, as it remains finite and interpretable even when the distributions have non-overlapping support.

In this study, the JSD is employed to rigorously evaluate the similarity between the real pediatric diabetes dataset and the synthetic datasets generated by the GAN and WGAN models. The JSD is chosen because it provides a comprehensive measure of distributional similarity, capturing both general alignment and specific differences between the real and synthetic data distributions. This is particularly important given the complex nature of pediatric diabetes data, which may contain nuanced patterns and variability that must be accurately reflected in the synthetic data.

The implementation of the Jensen-Shannon Divergence (JSD) in this study is carried out through the following steps:

The use of JSD in this study is justified by its capacity to provide a clear and quantitative assessment of the extent to which the synthetic data replicates the distributional characteristics of the original data. By evaluating both global and feature-specific similarities, the JSD offers a robust measure of the performance of the generative models, ensuring that the synthetic datasets produced are suitable for further analysis and modeling within the context of pediatric diabetes research.

Kullback-Leibler Divergence (KLD)

The Kullback-Leibler Divergence (KLD) is a fundamental measure from information theory that quantifies the difference between two probability distributions^[17]. It is commonly used to evaluate the performance of generative models by measuring the "cost" or "loss of information" when approximating one probability distribution with another^[36][46]. In this study, KLD is employed to assess how closely the synthetic datasets generated by the GAN and WGAN models match the distribution of the real pediatric diabetes data.

Mathematically, KLD between two discrete probability distributions P (real data) and Q (synthetic data) is defined as shown in Equation (8):

where:

In this study, both JSD and KLD are used to provide a comprehensive evaluation of the synthetic data quality. The combination of these two metrics offers a robust framework for assessing how effectively the generative models replicate the distributional characteristics of the real pediatric diabetes data. While JSD provides a global measure of similarity that is symmetric and bounded, making it suitable for general assessments of distributional overlap, KLD offers a more detailed perspective by identifying specific divergences where the synthetic data may differ significantly from the real data^[47][48].

The implementation of KLD follows these steps:

The use of both JSD and KLD in this study is justified by their complementary strengths. JSD is robust to extreme discrepancies and provides an overall measure of similarity, making it useful for confirming that the synthetic data generally aligns with the real data. KLD, on the other hand, offers a more granular assessment, pinpointing exact areas of divergence, which is crucial for refining the generative models. Together, these metrics ensure that the synthetic datasets generated are not only globally similar to the real data but also accurately capture its detailed characteristics, thereby validating their use for subsequent analysis and predictive modeling in pediatric diabetes research.

Predictive Model Evaluation

The quality of synthetic datasets is not only defined by their statistical alignment with real data but also by their practical utility in downstream machine learning tasks. To this end, a predictive model evaluation framework is incorporated as an additional metric for assessing the efficacy of the generated synthetic data. This method evaluates how well synthetic datasets support classification tasks, serving as a proxy for their utility in real-world applications.

The Random Forest algorithm is selected as the predictive model due to its inherent robustness, capacity to handle high-dimensional data, and ability to model complex feature interactions^[49][50]. Random Forest operates as an ensemble of decision trees, where each tree is constructed using a bootstrap sample of the training data. At each decision node, a random subset of features is considered for splitting, introducing variability that enhances the generalizability of the model. The final prediction is obtained through majority voting (classification) or averaging (regression) across all trees, ensuring resilience to overfitting^[51].

Mathematically, the Random Forest Classifier constructs T decision trees, where each tree t ϵ T minimizes the Gini impurity Ig for classification tasks, as shown in Equation (9).

Here, C denotes the number of classes, and p_i , is the proportion of samples belonging to class i in the current node. This criterion guides the selection of optimal splits at each node, ensuring that the trees effectively capture the underlying patterns in the data^[52].

By employing Random Forest, this study introduces a pragmatic approach to assess the synthetic data's relevance in predictive modeling. This framework complements statistical measures like Jensen-Shannon Divergence (JSD) and Kullback-Leibler Divergence (KLD) by providing a task-oriented perspective, ensuring a comprehensive evaluation of synthetic data quality.

Feature filtering

The preprocessing of the dataset involves several critical steps to ensure its suitability for synthetic data generation and subsequent analysis. The original dataset comprises 834 records with 22 features, including various demographic, clinical, and biochemical variables relevant to pediatric diabetes research. To improve the dataset's quality and focus, three features are removed due to their lack of direct relevance to the study objectives, resulting in a refined dataset containing 19 features. The excluded features are detailed in Table 4.

Data Incompleteness

Handling data incompleteness presents a substantial challenge in the preprocessing stage. Among the 834 records, 462 (approximately 55.4 %) exhibit at least one missing value. To address this issue, an imputation strategy is employed, wherein missing values are replaced with the mean of their respective features. This approach maintains the dataset's integrity by preserving the total number of records and minimizing potential biases that could arise from data omission. Following this imputation process, the dataset retains all 834 records, as summarized in Table 5.

These preprocessing steps are vital in preparing the dataset for further analysis. They ensure that the data is clean, complete, and properly scaled, thereby enhancing the reliability and quality of the synthetic data generated.

Data normalization

After the imputation of missing values, a normalization process is applied to ensure consistency in the scale of the dataset's features, which is essential for the effective training of generative models. A Min-Max normalization technique is utilized, scaling each feature to a range between 0 and 1. This standardization mitigates the risk of features with larger numerical ranges disproportionately influencing the model training process, thereby ensuring that all features contribute equitably to the generative modeling efforts.

Evaluation of GAN-Generated Synthetic Data

The Generative Adversarial Network (GAN) is utilized to create a synthetic dataset designed to closely replicate the distributional characteristics of the original pediatric diabetes data. A synthetic dataset consisting of 417 records, representing 50 % of the original dataset's size, is generated to maintain a balance between capturing the complexity of the data and ensuring a robust comparison between the real and synthetic datasets.

The effectiveness of the GAN in generating realistic synthetic data is evaluated through a comparative analysis of the global distributions of the real and GAN-generated data. Figure 3 illustrates this comparison by aggregating the normalized values of all features into a single composite measure for each record, providing a comprehensive view of the similarity between the two datasets. This composite measure summarizes the overall alignment between the datasets, capturing both the central tendencies and variability of the original data.

Figure 3 Global Distribution Comparison Between Real and GAN-Generated Data. 

In Figure 3, the "Aggregated Feature Value" represents a combined metric derived by summing all normalized feature values for each record, while the probability density indicates the relative frequency of these aggregated values across the dataset. The curve corresponding to the real dataset is depicted in blue, while the red curve represents the GAN-generated dataset. The visualization demonstrates that the synthetic data generated by the GAN aligns well with the general shape of the real data distribution, capturing key patterns and trends, particularly in the central regions where the curves overlap closely.

Discrepancies are observed in the tails of the distribution, where the GAN fails to replicate less frequent and extreme values. This limitation arises from the training instability commonly associated with traditional GAN architectures, particularly when modeling complex, high-dimensional data. While the GAN successfully reproduces the dominant statistical features of the real dataset, these deviations highlight its reduced ability to capture rare patterns that require more stable and fine-tuned training processes.

To further quantify the similarity between the real and GAN-generated data, two divergence metrics are employed: the Jensen-Shannon Divergence (JSD) and the Kullback-Leibler Divergence (KLD). The values of these metrics, presented in Table 6, offer a detailed view of the effectiveness of the GAN in approximating the real data.

As shown in Table 6, the Jensen-Shannon Divergence (JSD) value of 0.482 indicates a substantial degree of similarity between the distributions of the real and GAN-generated data. The relatively low JSD suggests that the GAN effectively captures the overall structure and dominant statistical features of the original dataset. This result confirms that the global patterns, including central tendencies and frequent values, are well-represented in the synthetic data.

The Kullback-Leibler Divergence (KLD) value of 1.617 provides additional insight into the alignment of the two distributions. While this value remains relatively low, it is slightly higher than the JSD, indicating that discrepancies are more pronounced in regions with lower probabilities, such as the tails of the distribution. This reflects the GAN's difficulty in modeling less frequent or extreme values, which aligns with the limitations observed visually in Figure 3. Together, these divergence metrics demonstrate that while the GAN-generated data closely approximates the real data, there is room for improvement in capturing rare patterns and achieving a more precise representation of the complete data distribution.

To complement this evaluation, the quality of the GAN-generated synthetic data is also assessed through its utility in a supervised learning context. A Random Forest classifier is trained separately on the real dataset and the GAN-generated synthetic dataset, using a binary classification task where the target variable indicates the presence or absence of diabetes. The feature Insulin (labeled as Insulin_ugml in the dataset) is used to define this target variable, as insulin levels are clinically recognized as a key indicator for diabetes diagnosis. Based on medical thresholds, records with insulin levels greater than 25 μU/mL are labeled as diabetic (1), while others are labeled as non-diabetic (0).

Key metrics, including accuracy, precision, recall, F1-score, and ROC AUC, are computed to evaluate the performance of the model in each case. The results, summarized in Table 7, provide a point of comparison to understand how well the synthetic data supports predictive modeling relative to the real data.

The performance metrics in Table 7 show that the GAN-generated synthetic data achieves strong and consistent results across all evaluated metrics. Although a slight drop of approximately 4 % is observed in accuracy, precision, recall, and F1-score compared to the real data, this difference is within acceptable limits for synthetic datasets. The marginal decline reflects the challenge of fully replicating the nuanced variability of real-world data, particularly in capturing less frequent patterns or edge cases.

The relatively high precision and recall values indicate that the synthetic data retains the critical predictive features necessary for identifying positive cases while effectively minimizing false positives. This balance is essential for clinical applications, where both sensitivity (recall) and specificity (precision) are critical for reliable predictions. Furthermore, the high ROC AUC score confirms that the model trained on GAN-generated data maintains a strong discriminative capability, demonstrating its ability to distinguish between classes effectively.

These findings highlight that, despite minor discrepancies, the GAN-generated data successfully preserves the underlying relationships and predictive features of the real dataset. This validates the GAN's practical utility for training machine learning models in scenarios where real data are scarce or access is restricted, offering a robust alternative for enhancing predictive modeling in pediatric diabetes research.

However, while these results are promising, they also reflect the inherent challenges of working with pediatric data, which tend to be highly variable due to the ongoing growth and development of children. The diversity in growth rates, physiological changes, and individual responses to health conditions can result in complex patterns that are difficult to replicate fully with synthetic models. These characteristics contribute to the observed differences between the real and synthetic data and underscore the need for careful consideration when generating synthetic datasets in this context.

Overall, the combined analysis of the visual comparison, divergence metrics, and supervised learning performance suggests that the GAN is highly effective in generating synthetic data that not only closely matches the general distribution of the real data but also retains the critical predictive features necessary for accurate classification. The comparable performance of the Random Forest classifier on both the real and synthetic datasets reinforces the validity of the synthetic data, demonstrating that it encapsulates the essential patterns and relationships required for supervised learning tasks. This comprehensive evaluation highlights the potential of GAN-generated data to serve as a reliable alternative for training machine learning models when real data is scarce or inaccessible.

Evaluation of WGAN-Generated Synthetic Data

The Wasserstein Generative Adversarial Network (WGAN) is the second method used for generating synthetic data in this study, aimed at producing a dataset that closely replicates the characteristics of the original pediatric diabetes data. The WGAN model generates a synthetic dataset of 417 records, mirroring the size used in the GAN-generated dataset to maintain consistency in comparisons.

The performance of the WGAN is evaluated by comparing the global distributions of the real and WGAN-generated data, as shown in Figure 4. As in the GAN analysis, the normalized values of all features are aggregated into a single composite measure for each record, allowing a direct comparison of the overall alignment between the datasets. This method provides a comprehensive view of how well the WGAN captures the underlying structure and variability of the real data.

Figure 4 Global Distribution Comparison Between Real and WGAN-Generated Data. 

In Figure 4, the "Aggregated Feature Value" represents a combined metric derived by summing all normalized feature values for each record, while the probability density indicates the relative frequency of these aggregated values across the dataset. The curve corresponding to the real dataset is depicted in blue, while the red curve represents the WGAN-generated dataset. The visualization shows that the WGAN-generated data aligns closely with the real data distribution, particularly around the central peak and mid-range values, where the curves exhibit almost perfect overlap. This alignment indicates that the WGAN successfully captures the most frequent and dominant statistical patterns in the data, closely approximating the global structure observed in the real dataset.

Compared to the traditional GAN, the WGAN demonstrates a significant reduction in discrepancies at the tails of the distribution, where less frequent and extreme values occur. This improvement reflects the enhanced stability of the WGAN training process, which allows it to model complex patterns with greater accuracy. By reducing deviations in the tails, the WGAN effectively retains both dominant and rare features, preserving the variability essential for realistic synthetic data generation. This capacity makes the WGAN-generated data not only a reliable representation of the real dataset but also a valuable resource for downstream machine learning tasks, particularly in scenarios with limited real data availability.

To further quantify the similarity between the real and WGAN-generated data, the same divergence metrics used for the GAN evaluation, the Jensen-Shannon Divergence (JSD) and the Kullback-Leibler Divergence (KLD), are calculated. The results, presented in Table 8, provide a detailed assessment of the WGAN's performance.

As shown in Table 8, the Jensen-Shannon Divergence (JSD) value of 0.480 is slightly lower than the value obtained with the GAN (0.482), indicating a marginal improvement in the similarity between the WGAN-generated and real data distributions. This lower JSD suggests that the WGAN more effectively captures the global statistical structure of the dataset, aligning closely with the dominant patterns observed in the real data.

The Kullback-Leibler Divergence (KLD) value of 1.584 is also marginally reduced compared to the GAN (1.617), highlighting a better representation of less frequent and extreme values in the tails of the distribution. This improvement reflects the enhanced stability of the WGAN training process, which enables it to model complex, low-probability patterns with greater accuracy. Collectively, these results demonstrate that the WGAN outperforms the traditional GAN in both global alignment and in capturing subtle variations, making it a more reliable approach for generating high-fidelity synthetic data.

Similar to the evaluation conducted for GAN, the quality of the WGAN-generated synthetic data is assessed through its utility in a supervised learning context. This analysis follows the same approach as the previous section, where a Random Forest classifier is trained on both the real dataset and the WGAN-generated synthetic dataset to predict the presence or absence of diabetes based on insulin levels. Key metrics such as accuracy, precision, recall, F1-score, and ROC AUC are calculated to evaluate the model's performance on both datasets. The results are summarized in Table 9.

As shown in Table 9, the Random Forest classifier trained on WGAN-generated synthetic data achieves performance metrics that are closely aligned with those obtained using real data. The accuracy, precision, recall, F1-score, and ROC AUC for the WGAN-generated data show only a slight decrease of approximately 2 % compared to the real dataset. This minor drop highlights the ability of the WGAN to generate high-fidelity synthetic data that preserves critical predictive features and relationships.

When compared to the GAN-generated data (Table 7), the WGAN demonstrates improved performance across all metrics, particularly in recall and ROC AUC. The slight enhancements in recall indicate that the WGAN-generated data better captures edge cases and less frequent patterns, reducing false negatives more effectively. Additionally, the ROC AUC value of 0.97 for the WGAN surpasses the 0.94 obtained with the GAN, showcasing the superior discriminative power of the WGAN-generated data.

The WGAN’s ability to produce synthetic data with enhanced stability and improved representation of subtle and less frequent patterns demonstrates its superiority over traditional GANs. By addressing the limitations observed in earlier models, the WGAN-generated data offer a reliable and realistic alternative for training machine learning models. This is particularly valuable in pediatric research, where access to real-world datasets is often limited, and the quality of synthetic data plays a crucial role in enabling robust predictive modeling.

The results obtained in this study, particularly the superior performance of Wasserstein Generative Adversarial Networks (WGANs) compared to traditional GANs, align with findings from other research on synthetic data generation in healthcare. Goncalves et al.^[55] conducted a thorough examination of synthetic data generation techniques for cancer patient data, focusing on methods such as probabilistic models, Bayesian networks, and GANs. Their results demonstrated the potential of these techniques to approximate real datasets, achieving a Kullback-Leibler (KL) divergence score of 0.47 for cancer datasets, which is comparable to the results obtained in this study for pediatric diabetes data (KL divergence of 1.584 for WGAN-generated data). Although the divergence in this study is slightly higher, it is crucial to consider the unique challenges posed by pediatric data. Growth and developmental variability introduce significant complexity in modeling, making pediatric data inherently more difficult to capture accurately.

Other studies in synthetic healthcare data generation also reflect similar trends, particularly in the challenges of handling complex datasets. Xu et al.^[56] applied Conditional GANs (CTGAN) to generate synthetic tabular data and found that CTGAN outperforms Bayesian networks and other GAN-based models on several real-world datasets. Although their focus was not on pediatric data, their results highlight the effectiveness of using conditional GANs for generating structured data patterns. However, their work, like many others, focuses primarily on adult datasets, where the variability present in pediatric populations is not a significant factor. This difference underscores the need for models capable of adapting to more dynamic physiological conditions, such as those found in children with diabetes, where growth and development create a more challenging environment for accurate synthetic data generation.

While the generation of synthetic healthcare data has advanced significantly, much of the focus has been on adult datasets, with limited attention paid to pediatric populations. This study addresses a significant gap in the literature by applying advanced generative models, specifically GANs and WGANs, to pediatric diabetes data. Pediatric datasets present unique challenges due to the high variability in physiological characteristics as children grow and develop. This inherent variability complicates the modeling process and requires more sophisticated approaches to accurately reflect the underlying data distributions. To address these complexities, WGANs were employed due to their stability in training and ability to handle the nuanced data patterns found in pediatric populations. Moreover, the use of Jensen-Shannon Divergence (JSD) and Kullback-Leibler Divergence (KLD) as evaluation metrics ensures that the generated datasets maintain the essential statistical properties necessary for predictive modeling.

To complement the evaluation of these divergence metrics, the quality of the WGAN-generated synthetic data was further assessed through its utility in a supervised learning context. A Random Forest classifier, trained on both real and WGAN-generated synthetic data, revealed that the synthetic data maintained essential predictive features, with classification performance metrics (accuracy, precision, recall, F1 score, and ROC AUC) comparable to the real data. These results demonstrate that the WGAN-generated data can be effectively used in predictive models for pediatric diabetes, showing that it not only captures the global distribution of the real data but also retains the critical features required for accurate classification.

By generating high-quality synthetic datasets for pediatric diabetes, this work can significantly enhance predictive models and personalized treatment approaches, addressing the critical scarcity of real-world pediatric datasets in this domain. To our knowledge, no studies to date have applied generative models specifically to pediatric data in the context of diabetes research, making this contribution particularly novel.

Limitations and adaptability

While this study demonstrates the successful use of GANs and WGANs to generate high-quality synthetic datasets for pediatric diabetes research, certain limitations must be acknowledged. The quality and representativeness of the generated synthetic data depend heavily on the characteristics of the input dataset. Small sample sizes, incomplete data, or datasets with imbalanced classes may hinder the models’ ability to capture rare or subtle patterns accurately. This limitation is particularly significant in pediatric datasets, where inherent variability due to growth, development, and physiological changes introduces additional challenges. Ensuring high-quality input data remains a critical prerequisite for achieving reliable synthetic data generation.

Another limitation relates to the specificity of the current methodology to pediatric diabetes data. Although GANs and WGANs are generalizable techniques, their application to other datasets or domains requires adaptation. For instance, datasets with different structures, such as high-dimensional multi-omics data or longitudinal patient records, may require modifications to the preprocessing pipeline, including tailored strategies for handling missing data, normalization techniques, or feature selection. Likewise, adjustments to the model architectures may be necessary to optimize performance for datasets with greater complexity or variability.

Despite these limitations, the proposed methodology provides a solid foundation for generating synthetic datasets in healthcare and beyond. Future research can explore its adaptability to other medical conditions, larger datasets, and alternative domains, such as genomics or environmental monitoring, where data scarcity is also a significant challenge. Additionally, further evaluation across diverse datasets could validate the generalizability of this approach, ensuring its broader applicability in machine learning-driven predictive modeling.