Type 1 diabetes (T1D) is an autoimmune condition in which the body’s immune system selectively destroys pancreatic beta cells, resulting in absolute insulin deficiency and lifelong dependence on exogenous insulin. It primarily affects children and adolescents and requires continuous insulin administration, glucose monitoring, precise dietary management, and lifestyle adjustments [1]. Recent data from the International Diabetes Federation Diabetes Atlas 2025 indicate a significant rise in the global burden of T1D, with an estimate of 30,000 children and youth worldwide at risk of death due to undiagnosed T1D at onset [2]. In 2024, more than 9.5 million people were living with T1D globally, including approximately 1.9 million children and adolescents [2]. Within this context, Saudi Arabia is among the countries most affected globally, with 46,469 children and adolescents reported to be living with T1D in 2024 [2]. This burden is further compounded during adolescence, when hormonal and psychological changes can affect treatment adherence and increase the risk of complications [3,4].

Acute complications, such as diabetic ketoacidosis (DKA) and severe hypoglycemia, are among the leading causes of mortality in this age group if not promptly addressed. In many patients, approximately 30% of first hospital admissions and initial diagnoses of T1D occur due to DKA [1,4]. DKA occurs as a result of absolute insulin deficiency, which increases lipolysis, leading to uncontrolled hyperglycemia, ketone body production, and metabolic acidosis, and may be fatal if not treated promptly [3,4]. Severe hypoglycemia, on the other hand, results from an imbalance between insulin levels and glucose availability, which can cause seizures, loss of consciousness, or sudden deterioration [4]. Both acute complications represent life-threatening emergencies and remain a significant challenge contributing to mortality in children and adolescents with T1D [1].

In addition to acute risks, long-term poorly controlled T1D leads to chronic complications, including nephropathy, chronic kidney disease (CKD), and neuropathy [3]. These complications may begin during adolescence and progress over time, negatively impacting quality of life and increasing health care and economic burdens [3,5]. Clinical evidence underscores that maintaining blood glucose levels near normal, as monitored by hemoglobin A_1c (HbA_1c), significantly reduces the long-term incidence of these adverse outcomes [3,6]. This highlights the importance of early risk prediction and advanced therapeutic approaches to mitigate long-term adverse outcomes [1,5].

The medical literature emphasizes that early intervention in chronic conditions, such as T1D, can lead to significantly better outcomes. Accurate risk prediction supports more timely preventive care, reduces hospital admissions, improves patients’ daily lives, helps health care providers allocate resources effectively, and also strengthens patient and family education [3-5].

Despite the abundance of available clinical data, a noticeable gap remains in the availability of interpretable tools that assist clinicians in reliable and meaningful risk assessment, particularly for pediatric populations [1,5].

Machine-learning (ML) studies on predicting diabetes complications have made notable advances. Recent reviews indicate significant advancements in ML-based detection, but persistent challenges remain in translating these models into clinical practice, particularly in integrating them into established medical workflows [7]. Most work has focused on the adult population with type 2 diabetes, with limited attention on children and adolescents with T1D. This focus leaves a gap in pediatric T1D care, where comprehensive multicomplication risk assessment remains limited [8].

Jian et al [5] developed predictive models that achieved high accuracy using common algorithms, such as random forests and decision trees. However, their work focused on adults and chronic complications without leveraging clinical guidelines. Similarly, Ravaut et al [9] applied gradient-boosted models to a large administrative dataset to predict both acute and chronic complications, although the binary outcome approach limits its use for nuanced risk stratification.

Eid et al [10] and Subramanian et al [11] focused on acute complications, such as DKA in pediatric patients, but their models were confined to specific outcomes without broader complication profiling or integration of clinical knowledge. Voskergian et al [12] employed synthetic electronic health records (EHRs) to predict multiple complications while not incorporating guideline-based features or ensuring interpretability.

Deep learning techniques, including deep neural networks, convolutional neural networks, and recurrent neural networks, demonstrate high predictive performance across various medical datasets. However, their lack of interpretability often makes them challenging to adopt in clinical practice [13]. By contrast, ML techniques, such as decision trees and logistic regression, are easier to interpret, making them more appropriate for use in pediatric health care contexts [10,14].

Although interpretation tools, such as SHAP (Shapley Additive Explanations), have improved model explainability [9], many existing studies still depend on basic importance scores or filter-based feature selection techniques and seldom contain domain-specific clinical insights [8,15]. This persistent gap highlights the growing need for approaches that provide transparent explanations, as emphasized by Netayawijit et al [16], while also integrating clinical expertise directly into the model’s logic to ensure relevance [8].

To bridge this gap, this study proposes an interpretable model specifically designed to predict T1D complications in children and adolescents. The model is developed using rule-based features extracted directly from the Saudi Diabetes Clinical Practice Guidelines (SDCPG), and it leverages SHAP and exhaustive feature selection (EFS) alongside a decision tree model to achieve interpretable and accurate predictions.

The primary aim of this study is to develop an interpretable predictive model designed to classify complication risk levels—low, chronic, and acute—among children and adolescents with T1D. By integrating expert clinical rules from the SDCPG with advanced feature selection methods (SHAP and EFS), this study bridges the gap between high predictive accuracy and clinical transparency. Ultimately, the study provides a locally aligned tool for the Saudi health care system that supports the P4 medicine (predictive, preventive, personalized, and participatory medicine) framework by offering proactive, evidence-based decision support for diabetes management.

The methodology aims to develop a multilevel predictive model that categorizes individuals with T1D based on their risk of complications. This classification is built on rules from the SDCPG. The model classifies patients into 3 main categories: acute risk, denoted by critical conditions, such as hypoglycemia and DKA; chronic risk, involving long-term complications, such as foot deformities, CKD, and neuropathy; and low risk, for patients showing no significant warning signs. By integrating expert clinical knowledge with data-driven modeling, the methodology seeks to identify the most consequential features in this classification. The process is structured into 5 main steps: data collection, data preprocessing, feature selection, training and validating the model, and finally evaluating different models to choose the most accurate one, as shown in Figure 1.

This study utilizes an open-source dataset titled “Dataset on Significant Risk Factors for Type 1 Diabetes,” published in 2018 [17]. The dataset targets children and adolescents in Bangladesh and is designed to explore the major risk factors associated with T1D. It includes 306 participants, with an equal distribution between those diagnosed with T1D and those without, collected through structured surveys from hospitals and diagnostic centers in Dhaka.

The dataset consists of 22 features covering a broad range of information—from demographic characteristics to clinical indicators, such as HbA_1c and hypoglycemia, in addition to lifestyle factors, comorbidities, and family history of diabetes. The types of data include categorical, temporal, continuous, and multilabel text variables. Table 1 presents a summary of the dataset features and their corresponding data types. Detailed feature descriptions and categorized values are provided in Multimedia Appendix 1.

This study involved a secondary analysis of a publicly available and anonymized dataset published by Asaduzzaman et al [17] (2018). The dataset contains deidentified records and does not include any personally identifiable information. As no new data were collected and no direct interaction with human participants occurred, institutional review board approval was not required for this analysis. Informed consent was not required, as the dataset is publicly available and contains no identifiable information. All data were handled in accordance with relevant ethical standards, ensuring the privacy and confidentiality of individuals. No compensation was provided, as this study did not involve direct participant recruitment.

The dataset was split into 2 subsets: 80% used for training and 20% for testing. This allowed the model to learn from most of the data and check how well it performs on new, unseen examples. To ensure the class balance remained consistent in both sets, a stratified split was applied. This helped maintain the distribution of the target variable fairly and reduced bias in the results.

To comprehensively evaluate model performance, assessments were conducted in 3 stages: the training set, the test set, and CV to measure consistency and robustness. The evaluation metrics included accuracy, weighted precision, weighted recall, weighted F₁-score, and multiclass area under the curve (AUC). The evaluation focused on average performance and stability, using the mean and SD across folds.

Models were ranked based on their ability to balance performance metrics on the test set, with priority given to strong F₁ and AUC scores combined with low variability. Finally, the decision path of the highest-performing decision tree model was visualized to demonstrate the interpretability of its decision-making process.

The classification of patients into low, chronic, and acute risk categories was guided by clinical rules derived from the SDCPG. The class distribution shows that acute risk constituted the largest group (110/306 patients, 35.9%), followed closely by low risk (101/306 patients, 33%) and chronic risk (95/306 patients, 31%). This fairly balanced distribution enabled model training without requiring specialized class-balancing strategies. The slight increase in acute risk patients is expected, given the age group studied, as children and adolescents are more likely to experience sudden complications, such as hypoglycemia or DKA. In contrast, chronic conditions usually take longer to develop and are less common at younger ages [3].

Table 4 presents a detailed comparison of 5 predictive models developed using the top-performing feature subsets. Evaluations were conducted across training, testing, and CV phases to assess their accuracy and consistency. All models achieved a test F₁-score of 0.983, with a CV F₁ SD of 0.0189, reflecting strong and consistent performance. Despite performing similarly, all the models shared a common group of 5 core features: BMI, hypoglycemia, disease duration, HbA_1c, and impaired glucose metabolism. Variations between the models were limited to 1 additional feature per model, but these differences did not result in noticeable performance improvements.

The complete evaluation results, including all 1013 feature subsets tested via EFS, and detailed metrics for the top 5 models, are provided in Multimedia Appendices 4 and 5.

Although all 5 models demonstrated similar quantitative performance, the first model was selected as the final model due to its structural simplicity. It relies on 5 core features instead of 6 while still maintaining high predictive accuracy. Notably, these 5 features—hypoglycemic episodes, HbA_1c, BMI, disease duration, and impaired glucose metabolism—were consistently present across all top-performing feature sets. Figure 4 illustrates how the final model makes decisions through a simplified tree that reflects a clinically logical sequence. It begins by checking for hypoglycemic episodes, then evaluates HbA_1c if no episodes are recorded. For patients with hypoglycemia, it moves through BMI, disease duration, and impaired glucose metabolism sequentially to reach a final risk classification.

To demonstrate how the model processes patient information, Table 5 presents 2 hypothetical patients based on the decision paths shown in Figure 4.

For patient A, the model first checks for hypoglycemia. Because hypoglycemia is present, the model follows the right branch of the tree and evaluates HbA_1c. Since the HbA_1c level is below 7.5%, the patient is classified as an acute risk. For patient B, no history of hypoglycemia is present. The model follows the left branch of the tree and evaluates BMI and metabolic indicators. Because the patient is obese and shows impaired glucose metabolism, the model classifies the patient as a chronic risk.

To evaluate the contribution of this study, a comparative analysis was conducted with 5 representative ML studies targeting the prediction of diabetes complications. The goal of this comparison is to evaluate how this study differs in methodology, clinical integration, model interpretability, and performance outcomes. Table 6 presents general information on the selected studies, including their primary objectives, targeted complication types, dataset origin and size, and whether they incorporated formal clinical guidelines. Table 7 summarizes the modeling approaches used in each study, reported performance metrics, and levels of interpretability. Including this study, a total of 6 models are summarized.

A key differentiator of this work is its ability to classify both acute and chronic complication risks through a structured 3-level classification (low, chronic, and acute), whereas most prior studies focused on predicting only a single complication type, typically either acute (eg, DKA) or chronic (eg, CKD). While Ravaut et al [9] addressed both complication types, their model still adopted a binary outcome structure, lacking the nuanced stratification offered in this study. Additionally, this study is the only one among the reviewed works to utilize national clinical guidelines (SDCPG) to inform risk labeling, thereby strengthening its clinical alignment.

Despite using a relatively small open-source dataset (N=306), the proposed model achieved a test F₁-score of 0.98 and AUC ≈ 0.98, on par with or exceeding the performance of more complex models built on larger datasets. Furthermore, it employs only 5 clinically meaningful features and leverages a decision tree classifier enhanced by the SHAP analysis, providing both transparency and clinical explainability. This comparison emphasizes that high predictive accuracy can be achieved without sacrificing interpretability, especially when models are designed with clinical context and usability in mind.

The model demonstrated powerful performance while maintaining a clear and interpretable structure, which makes it a practical choice for clinical use. In addition to its technical strengths, this model, based on SDCPG, ensures consistency with local clinical practice. This alignment enhances its credibility and supports smooth integration into existing clinical systems, increasing its potential for real-world adoption. Its ability to identify complication risks in children and adolescents with T1D ensures early intervention and supports a shift toward preventive care rather than reactive treatment.

In this context, the model contributes specifically to the predictive principle of P4 medicine by enabling the early identification of complication risk in this population. While other principles of P4 medicine—such as preventive, personalized, and participatory strategies—require further integration, this model offers a foundational predictive tool to support future enhancements. It not only anticipates complications but also adapts to individual clinical profiles and offers transparent decision paths that can be shared with both patients and health care providers.

While the model yielded encouraging outcomes, several important limitations should be considered. The dataset used in this study was published in 2018 and was derived from a single center, with a relatively small sample of 306 pediatric patients, which may affect the model’s ability to generalize, especially in rare or borderline presentations. Since model validation was conducted without external validation, the findings may not fully translate to real-world settings, which could limit how well the model performs across broader populations.

A further limitation is that the risk labels were generated using SDCPG-derived rules rather than independently observed clinical outcomes, such as confirmed DKA or nephropathy. Therefore, the presented model performance reflects adherence to guideline-based classification logic, rather than prospective prediction of clinical complications. Future longitudinal investigations are required to validate the proposed risk classifications using real-world clinical outcomes.

Additionally, the clinical data were originated from a Bangladeshi cohort, whereas the risk classification rules were derived from the SDCPG. Although this supports guideline portability, differences in health care infrastructure, clinical practices, or population characteristics may influence generalizability. Some clinical features were simplified, such as using age groups instead of exact values, and certain variables were inconsistently recorded or structured, which may reduce predictive precision in pediatric patients. Moreover, the study did not compare its model against an established clinical risk scoring system or clinical decision support system baseline tool, as no standardized benchmark tool was available for this specific population and use case.

Finally, although the model uses common clinical features, such as HbA_1c, disease duration, and BMI, inconsistencies in data documentation and system integration across health care institutions could pose challenges for its direct implementation into clinical decision support system platforms.

Future work should aim to validate the model using external datasets from diverse populations to assess its generalizability. Additionally, exploring the integration of other P4 medicine elements, such as personalized treatment pathways and participatory tools, could enhance the model’s utility. Collaboration with health care institutions to embed the model into existing EHR systems and collect feedback from clinicians on usability will be critical for real-world applications and iterative refinement.

This study developed a clinically meaningful approach to classifying complication risks in children and adolescents with T1D, based on a set of clinical rules extracted from the SDCPG, to build a model that balances accuracy with interpretability. Using a hybrid feature selection technique that combines SHAP and EFS with a decision tree model, the model achieved consistently high performance using only 5 clinical indicators. These results suggest that effective risk classification can be achieved without complex systems, making the model a practical candidate for clinical use. Its transparency makes it easier for health care teams to trust and apply.

Future work should focus on external validation to test its generalizability and on expanding the dataset to improve robustness. Further exploration of the model’s integration within EHR systems and alignment with the broader principles of P4 medicine is also warranted. This study specifically addresses the predictive component, laying the foundation for future work that could incorporate preventive, personalized, and participatory dimensions to enhance diabetes care for children and adolescents.