This section addresses the critical, complex, and multifaceted aspects of multimodal data acquisition, encompassing the collection of diverse data types from various sources and the requisite, often intricate, data preparation processes. These preparation stages, including but not limited to data cleaning, normalization, transformation, and feature engineering, are indispensable for effectively fueling the advanced AI models that form the analytical core of this research, ensuring they operate on high-quality, meaningful inputs to generate reliable and actionable insights [29,30]. The overarching goal is to establish robust data and conceptual groundwork upon which sophisticated risk prediction and stratification can be built, ultimately aiming to transform pediatric health care through proactive and personalized interventions. Conceptually, the framework aims to map a complex set of input variables Xdata (representing multimodal pediatric data) through a series of transformations and AI modeling of fAI to an interpretable risk stratification outcome Yrisk :

This high-level representation underscores the journey from raw data to actionable risk assessment.

The creation of a robust pediatric health risk stratification system requires a meticulously defined conceptual framework that delineates the core components, their interactions, and the overall data flow. As depicted in Figure 1, the proposed AI-driven framework is architected to be modular, scalable, and adaptable to diverse pediatric health contexts. Its primary objective is to convert raw, multisource pediatric data into actionable risk profiles, thereby facilitating proactive and personalized health care interventions.

Commencing with multimodal data input, the framework assimilates a wide array of data sources pertinent to child health. Subsequently, the data preprocessing and harmonization layer comes into play. Given the heterogeneity and potential quality issues inherent in real-world pediatric data, such as missing values, inconsistencies, and varying formats from different sources, this layer uses data cleaning, normalization, transformation, and integration techniques to create a unified and analysis-ready dataset.

Next is the feature engineering and selection module, which is dedicated to extracting meaningful features from the preprocessed data that are highly indicative of pediatric health risks. This process may involve creating composite variables, transforming existing features, or applying dimensionality reduction techniques to optimize the input for AI models. Domain knowledge from pediatric medicine is crucial in guiding this process to ensure clinical relevance.

At the core of the framework lies the AI-powered risk modeling and stratification engine. This engine uses advanced ML algorithms, with a particular focus on ensemble learning techniques, as discussed in the Introduction, to construct predictive models for various pediatric health risks. Trained on historical data, these models learn complex patterns and relationships between input features and health outcomes or risk states. The output of this engine is not merely a binary prediction but a nuanced risk stratification, which may encompass risk scores, probability estimates, or the identification of distinct risk phenotypes. The determination of weights for each data modality is a critical aspect of our ensemble model. Rather than assigning predetermined, fixed weights, our stacking framework uses a data-driven approach.

The final layers of the framework are devoted to risk profile generation and visualization and clinical decision support and intervention pathways. The generated risk profiles are designed to be interpretable by health care professionals and understandable by caregivers, often incorporating visual aids and clear explanations. This output is intended to seamlessly integrate with clinical workflows, offering evidence-based insights to support shared decision-making regarding preventive strategies, further assessments, or targeted interventions. Continuous monitoring and model updating are also integral to the framework, ensuring that the AI models maintain their accuracy and relevance as new information emerges and clinical knowledge evolves [31].

The primary pediatric health domains identified for this framework, as detailed in Table 2, cover a broad spectrum of child well-being. Patient safety focuses on risks, such as adverse drug events and medical errors, typically informed by EHR data and incident reports. Continuous health monitoring is crucial, tracking growth parameters and developmental milestones using EHR data and wearable sensor inputs. For children with ongoing conditions, chronic disease management addresses factors like glycemic control or asthma exacerbation frequency, primarily using EHR data and home monitoring device information. Mental health is a significant domain, evaluating behavioral indicators and emotional well-being via parental, teacher, and direct child assessments. Furthermore, preventive care adherence is monitored through risk factors, such as vaccination status and well-child visit compliance, drawing from EHR data and parental questionnaires [32]. The interconnected domains of nutrition and growth consider risks associated with nutritional intake and growth percentiles, also informed by EHR and parental questionnaire data. Lastly, physical activity levels and sedentary behavior duration, identified as critical risk factors, are assessed using wearable sensor data and parental questionnaires.

For each identified health domain, a set of specific, measurable risk factors and associated potential data sources is cataloged. This systematic identification ensures that the subsequent data acquisition and AI modeling efforts are focused and aligned with the goal of comprehensive risk assessment. The dynamic nature of these risk factors across different developmental stages (infancy, early childhood, middle childhood, and adolescence) is also a key consideration in the framework’s design and application.

The effectiveness of the AI-driven risk stratification framework is fundamentally dependent on the availability of comprehensive, high-quality, multimodal data that accurately reflect the various factors influencing child health. This section details the strategies for acquiring and preparing such data for input into the framework, emphasizing the critical importance of ethical considerations, including informed consent, data privacy, and security, in full compliance with relevant regulations, such as the General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA), adapted for pediatric populations.

The data sources for this framework are inherently multimodal, encompassing a variety of information types. EHRs serve as a primary source, offering longitudinal clinical data that include diagnoses (ICD codes), procedures, medications, laboratory results, growth chart data, and clinical notes. Access to both structured and unstructured EHR data is essential. Parental and child questionnaires or surveys are used to capture information often not systematically recorded in EHRs, such as detailed family history, socioeconomic status, lifestyle factors (diet, physical activity, and sleep), environmental exposures, and patient-reported outcomes or symptoms. Wearable sensors and mHealth data provide opportunities for continuous, real-world monitoring of physiological parameters (eg, heart rate, activity levels, and sleep patterns) and behavioral data. School health records can contribute information on immunizations, health screenings conducted at school, and potentially attendance or behavioral notes relevant to health [33]. Public health and environmental databases can offer community-level data on factors like air quality, neighborhood socioeconomic indicators, and access to health care or recreational facilities. Genomic data, when applicable and ethically sourced, can provide valuable genetic risk scores or specific genetic markers associated with pediatric conditions.

The data preparation pipeline, illustrated in Figure 2, involves several key steps to transform raw, multimodal data into an analysis-ready dataset. Initially, data from disparate sources must undergo integration and harmonization. This process includes mapping data elements to common terminologies, resolving inconsistencies, and creating a unified data schema. Data cleaning is then performed to address missing values, correct errors, and remove outliers. Data transformation may be necessary to convert raw data into formats suitable for analysis, such as calculating age-specific z scores for growth parameters or deriving summary statistics from time-series sensor data. For instance, if xraw is a raw measurement and μage,sex and σage,sex are the age- and sex-specific mean and SD, respectively, from a reference population, the z score can be calculated as follows:

This normalization is crucial for enabling comparisons of measurements across different age groups and sexes. Finally, data anonymization or pseudonymization is rigorously applied to protect patient privacy before the data are used for model development. The resulting dataset provides the foundation for the feature engineering and AI modeling stages described in the subsequent part.(2)

The integrity and utility of input data are paramount to the performance and clinical relevance of any AI model, particularly in the pediatric domain, where data can be sparse, longitudinal, and highly variable due to growth and development. Pediatric health data, often derived from multiple heterogeneous sources, such as EHRs, parental questionnaires, school health notes, and wearable sensors, typically present unique challenges, including systematically missing values (eg, developmental assessments not performed at certain ages), measurement noise (eg, variability in home-based measurements), and the presence of irrelevant or redundant information that can obscure true risk signals. Therefore, rigorous data preprocessing and thoughtful, domain-informed feature engineering are indispensable steps to prepare a high-quality dataset for subsequent AI modeling.

The initial phase of preprocessing involves comprehensive data cleaning. Missing data represent a pervasive issue in longitudinal pediatric health care datasets. For instance, growth parameters or developmental screening results might be missing for specific well-child visits. While simple imputation techniques like mean, median, and mode imputation can be applied for variables with a low percentage of missingness and random patterns, they often fail to capture the underlying data structure in pediatric cohorts. More sophisticated methods are typically required, such as k-nearest neighbors imputation, which identifies the k most similar pediatric cases (based on a suite of other observed features) and imputes the missing value for a feature Xj using a weighted average or majority vote from the neighbors [34]. Alternatively, model-based imputation, using algorithms like multivariate imputation by chained equations (MICE), can be used. MICE iteratively models each variable with missing values as a function of other variables in the dataset, cycling through variables until convergence is achieved. For a set of variables X1,…,Xp, MICE specifies a conditional model PXj∣X-j,ϕj for each Xj and iteratively samples from these conditional distributions.

Outlier detection and appropriate treatment are also critical in pediatric data, where extreme values might represent genuine clinical concern or measurement error. Statistical methods, such as the z score and IQR, are used. For a data point xi in a feature distribution, its z score is calculated as zi=(xi−μ)/σ, where μ is the mean and σ is the SD. Data points with |zi|>θoutlier may be flagged. The IQR method defines outliers as points falling below Q1-1.5×IQR or above Q3+1.5×IQR, where Q1 and Q3 are the first and third quartiles, respectively, and IQR=Q3-Q1. Decisions on handling outliers (removal, transformation, or winsorization) are made cautiously, considering the potential clinical significance of extreme values in child health.

Feature engineering in the pediatric context is a highly domain-driven process focused on creating new, informative features from raw data to capture developmental trajectories, critical exposure periods, and clinically relevant interactions. This enhances model performance and interpretability. In child health, this includes deriving age- and sex-adjusted z scores for growth parameters, such as height, weight, BMI, and head circumference, based on standardized pediatric growth charts like World Health Organization (WHO) and Centers for Disease Control and Prevention (CDC) growth standards. It also involves creating features like developmental velocity, calculated as ΔD/Δt=Dt2-Dt1/t2-t1, reflecting the rate of change in developmental milestone scores between assessments. Additionally, it covers quantifying cumulative exposure to risk factors, such as days with poor air quality and screen time during early childhood, generating features for adherence to pediatric guidelines like vaccination completeness and well-child visit schedules, and creating interaction terms reflecting synergistic effects in child health, such as genetic predispositions interacting with environmental exposures.

To reduce dimensionality, mitigate the risk of overfitting (especially with potentially limited sample sizes in specific pediatric subpopulations), and improve model training efficiency and interpretability, various feature selection techniques are used. These are broadly categorized into filter methods, wrapper methods, and embedded methods. Filter methods evaluate features independently of the AI model, using statistical measures relevant to pediatric outcomes, such as the chi-square test for assessing associations between categorical risk factors (eg, maternal smoking during pregnancy) and a binary child health outcome, and the ANOVA F value for assessing the relationship between numerical predictors (eg, birth weight) and different risk groups. Wrapper methods, such as recursive feature elimination, use the performance of a specific AI model to iteratively build and evaluate models with different subsets of features. Embedded methods perform feature selection as an integral part of the model training process. For instance, tree-based algorithms like random forest and gradient boosting inherently provide feature importance scores based on how much each feature contributes to reducing impurity or error across the ensemble of trees. Regularization techniques, such as L1 regularization (Lasso), are particularly useful. The Lasso objective function for a linear model predicting a child’s health outcome y based on features X and coefficients β is as follows:

where λ is the regularization parameter that controls the penalty on the sum of absolute values of the coefficients, effectively shrinking less important feature coefficients to zero. The selection of λ is typically done via cross-validation.

To provide an overview of the preprocessing pipeline and feature engineering strategies across different data modalities, the key steps are summarized in Table 3.

The process of risk stratification involves translating the continuous risk probabilities or composite risk scores generated by AI models into distinct, actionable risk categories that are meaningful in a clinical context. This is achieved by defining specific thresholds that convert model outputs into discrete risk strata. The thresholds are determined through a combination of clinical expertise, statistical methods, and decision theory principles, ensuring they are both clinically relevant and statistically valid. P(Risk / Xchild) represents the feature vector for a specific child. For a 4-tier stratification system, the risk categories can be defined as follows:

The selection of the thresholds θ1,θ2,θ3 is guided by clinical expertise, statistical methods, and decision theory. Pediatric specialists define thresholds based on established clinical guidelines, acceptable risk levels for specific age groups, and the availability and efficacy of preventive interventions. Statistical methods, such as using percentiles of the predicted risk distribution in a well-characterized pediatric reference population, can also be used. For example, thresholds might be set to correspond to the 50th, 80th, and 95th percentiles of this distribution. Decision theory principles can further refine threshold optimization by considering a utility function that balances the costs and benefits associated with true positives, false positives, true negatives, and false negatives. This might involve analyzing metrics like Youden J statistic (J=sensitivity+specificity−1) or finding points on the precision-recall curve that correspond to desired tradeoffs for pediatric screening or intervention programs.

When the system is designed to predict multiple distinct health risks for a child, such as the risk of obesity, developmental delay, or asthma exacerbation, the framework requires a mechanism to present these risks in a consolidated and understandable manner. This can be achieved by either displaying the stratified risk level for each condition independently or by developing a composite pediatric health vulnerability index (CPHVI). The CPHVI is calculated by assigning weights wi to each risk score S(Ri) based on factors, such as the clinical severity of risk Ri, its impact on long-term child development, its prevalence in the target pediatric population, and its responsiveness to an intervention. To translate the CPHVI into clinical practice, we propose a tiered intervention model based on predefined, clinically meaningful thresholds developed in collaboration with pediatric experts. For example, a CPHVI score exceeding a “high risk” threshold could automatically trigger an alert in the EHR system for the primary care physician, along with a recommendation for a direct referral to a relevant specialist. The formula for the CPHVI is as follows:

The determination of the weights wi is a complex task, often requiring expert consensus or data-driven approaches. Another advanced approach involves using unsupervised learning techniques, such as clustering algorithms, applied to the vector of predicted individual risk probabilities for each child. This can help identify common co-occurring risk patterns or distinct pediatric subphenotypes that share similar multirisk profiles, guiding more tailored multifaceted interventions.

To enhance clinical utility and facilitate shared decision-making with families, the reasons behind a child’s stratification into a particular risk category must be as transparent and interpretable as possible. Techniques, such as SHAP and Local Interpretable Model-Agnostic Explanations (LIME), are used. For a given prediction f(Xchild), SHAP assigns an importance value (SHAP value, ϕj) to each input feature j, representing its marginal contribution to pushing the prediction away from a baseline. The sum of SHAP values for all features plus the baseline prediction equals the model’s output for that child:

This allows clinicians to understand which specific factors (eg, low birth weight, specific dietary patterns, and lack of physical activity) are most influential in determining a child’s assessed risk level, thereby guiding targeted advice and interventions.

BERT-based NLP is used to extract informative representations from unstructured pediatric clinical text (eg, consultation records, clinical notes, and parental narratives), which are subsequently used as input features for pediatric health risk stratification models.

Pediatric medical journals serve as a vital source of pediatric disease knowledge, encompassing a wealth of information on pediatric diseases, including clinical symptoms, diagnostic methods, treatment approaches, and prognoses. These journals are written by pediatric experts and researchers and are based on extensive clinical practice and scientific research. They contain detailed case studies, research findings, and expert opinions, providing a solid foundation for the extraction and application of pediatric disease knowledge. The knowledge derived from these journals can help health care professionals better understand the characteristics and progression of pediatric diseases, thereby improving the quality of pediatric health care services [14].

The pediatric disease knowledge base is a structured repository that integrates information from pediatric medical journals and other authoritative sources. It includes various aspects of pediatric diseases, such as disease names, symptoms, signs, laboratory test results, imaging findings, treatment methods, and prognoses. This knowledge base is designed to provide comprehensive and accurate information on pediatric diseases, supporting clinical decision-making and research. By organizing and structuring the information from pediatric medical journals, the pediatric disease knowledge base enables efficient retrieval and utilization of pediatric disease knowledge, facilitating the application of this knowledge in clinical practice and research.

The process begins with the ingestion of pediatric medical journal data into our system. These textual data, rich in pediatric disease knowledge, undergo meticulous preprocessing to align with BERT’s input requirements. The text is tokenized into a sequence of tokens [T1,T2,…,Tn], where each Ti represents a word or subword unit. Additionally, the pediatric disease knowledge base is incorporated to enhance the model’s understanding of pediatric medical concepts. BERT’s architecture consists of multiple transformer encoder layers, each equipped with self-attention mechanisms. The input tokens, enriched with special markers like [CLS] and [SEP], pass through these layers, generating a matrix of contextualized embeddings z={z1,z2,…,zn}. These embeddings capture intricate semantic relationships and contextual information within the text, forming the basis for subsequent knowledge extraction tasks.

To identify key pediatric disease entities and their relationships within the text, named entity recognition and relation extraction techniques are applied to the BERT-generated embeddings. For named entity recognition, a linear layer combined with a conditional random field is used. The conditional random field loss function is defined as follows:

where N is the number of samples and Pyi∣Xi is the probability of the true label sequence yi given the input sequence Xi. For relation extraction between entities, a classifier is trained using a cross-entropy loss function as follows:

where M represents the number of entity pairs, C is the number of relation types, and yi,j and yˆi,j are the true and predicted probabilities for relation j between the entity pair i, respectively. The extracted entities and relationships are then integrated into the pediatric disease knowledge base. This knowledge base, enriched with entities and relationships from pediatric medical literature, serves to enhance a medical chatbot’s responses. When users input symptoms, the chatbot leverages this knowledge base to generate accurate and contextually relevant diagnostic suggestions and risk assessments.

Figure 3 illustrates a comprehensive workflow where BERT’s robust language understanding capabilities are harnessed to extract valuable pediatric disease knowledge from medical texts. This knowledge is subsequently used to empower a medical chatbot, enabling it to deliver precise and informative responses to user inquiries regarding pediatric symptoms. The integration of BERT with pediatric medical journals and a pediatric disease knowledge base provides a powerful tool for advancing pediatric health care through improved diagnostic accuracy and personalized treatment recommendations.

Our knowledge extraction module is built upon BioBERT, a pretrained language model optimized for the biomedical domain. We chose BioBERT as our base model due to its demonstrated strong performance on various biomedical text mining tasks. For fine-tuning, we used a domain-specific corpus comprising over 50,000 articles from leading pediatric medical journals and over 100,000 deidentified clinical notes. The model was fine-tuned for 5 epochs with a learning rate of 2e⁻⁵. The performance of our fine-tuned model on key NLP tasks was rigorously evaluated on a manually annotated test set of 500 clinical notes, which were dual-annotated by pediatric domain experts to ensure quality.

The successful deployment, scalability, and usability of the AI-driven pediatric health risk stratification framework critically depend on a meticulously planned and well-engineered system architecture. This architecture must robustly support efficient multimodal data processing pipelines; reliable and timely execution of complex AI models; secure and compliant management of sensitive pediatric health data; and an intuitive, actionable interface for diverse end users, including pediatricians, specialist clinicians, public health officials, researchers, and potentially, with appropriate safeguards, parents or caregivers. A multitier, service-oriented architecture is proposed to ensure modularity, maintainability, and scalability. This typically comprises the following tiers:

Key functional system modules are designed to support the end-to-end pediatric risk stratification process (Figure 4).

The data flow within the system begins with secure data ingestion from multiple sources into the data tier. The application logic tier then orchestrates the preprocessing and feature engineering pipelines. The processed features are subsequently fed into the AI model execution engine for risk prediction. These predictions are passed to the risk stratification and profiling module, and the final, interpretable pediatric risk profiles are stored and made available for secure access and visualization through the presentation tier or via the API layer. The technology stack will be carefully chosen based on criteria, such as scalability for large pediatric populations, real-time performance requirements, security mandates for child data, ease of development and maintenance, and compatibility with existing health care IT infrastructure. This might include Python with libraries like Pandas, NumPy, Scikit-learn, TensorFlow/Keras, and PyTorch for AI or ML development; web frameworks, such as Django/Flask (Python), Spring Boot (Java), and Node.js, for the backend services; modern JavaScript frameworks like React, Angular, and Vue.js for building responsive and interactive frontend interfaces; and a combination of SQL and NoSQL databases as described for the data tier. Containerization technologies like Docker and orchestration tools like Kubernetes may be used for deployment and scaling.

The pediatric health risk stratification prototype system integrates multisource data input with advanced AI models to deliver a comprehensive risk assessment platform. As illustrated in Figure 5, the system collects diverse data types, including patient demographics, clinical records, and parent-reported outcomes, which serve as the foundation for subsequent analysis and risk assessment. This integration ensures a holistic view of pediatric health, enabling more accurate and nuanced risk predictions.

At the core of the system is a pretrained BERT model, specifically fine-tuned to handle pediatric health terminology and context. The model processes input text data through multiple Transformer modules, which encode the text to capture complex linguistic patterns and semantic relationships. Feature extraction layers then distil these encoded representations into risk-relevant features. These features are passed to risk assessment components where the actual risk stratification takes place, transforming raw data into actionable insights regarding children’s health risks.

The system’s implementation leverages a robust technology stack to ensure efficiency, scalability, and user-friendliness. The backend, developed using Python and its rich library ecosystem, handles data processing and model execution. It uses microservices, potentially orchestrated by lightweight frameworks like Flask, to manage distinct functionalities such as data ingestion, preprocessing, and model inference. Data storage follows a hybrid approach, combining PostgreSQL for structured data with MongoDB for unstructured or semistructured content. The frontend, built with modern JavaScript frameworks, offers tailored interfaces for different users, including physicians and parents. It includes features like risk profile visualizations and population-level risk trend dashboards. Docker is used for containerization to ensure smooth deployment and scalability of the application components.

To illustrate the practical utility of the framework, consider a hypothetical clinical scenario. Dr Smith, a pediatrician, begins her day by logging into the system’s dashboard. The system flags one of her patients, a 5-year-old child, for a significant increase in the composite risk score for developing obesity. Dr Smith clicks on the patient’s profile and is presented with an interactive, multimodal dashboard. An explainable AI feature, using SHAP values, highlights the primary contributing factors: a recent decrease in physical activity levels captured by wearable data and parent-reported dietary logs indicating high consumption of processed foods. Based on these insights, the system provides Dr Smith with evidence-based, actionable recommendations, including a referral to a pediatric nutritionist and a set of tailored educational materials for the parents. This scenario demonstrates how the framework can transform raw data into clinically actionable insights to facilitate timely and personalized preventive care.

The primary dataset used for training and validating the AI models included data from over 40,000 pediatric participants, with an age range of 2 to 8 years at the time of data collection or follow-up. The dataset was partitioned into 3 mutually exclusive sets: a training set (70% of the data), a validation set (15%), and an independent holdout test set (15%). To ensure temporal validity if longitudinal data were used, the split was performed such that data from earlier time periods were used for training and validation, while data from later periods were reserved for testing, mimicking a prospective evaluation.

To benchmark the performance of the proposed AI-driven framework, a selection of robust baseline models was implemented and rigorously evaluated on an identical dataset and predictive task. These comparators included established traditional statistical models frequently used in pediatric risk prediction, specifically logistic regression for binary risk outcomes, where the probability of risk P(Y=1∣X) is modeled as P(Y=1∣X)=(1+e−(β0+∑i=1p  βiXi))−1, and Cox proportional hazards models in scenarios where time-to-event data for risk onset were available. Furthermore, simpler yet effective ML algorithms, namely support vector machine (SVM), configured with various kernels (linear and radial basis function), and standard single decision trees, were also incorporated as baselines. In instances where existing, validated pediatric risk scores or established rule-based systems were pertinent to the specific health outcomes under investigation and applicable to the dataset, these were also included in the comparative analysis to provide a comprehensive performance context. To rigorously assess the statistical significance of our model’s superior performance, we used the DeLong test to compare the area under the receiver operating characteristic curve (AUC-ROC) of our proposed model with that of each baseline model.

This study involved research with human participants. The study protocol was reviewed and approved by the Institutional Review Board and Research Ethics Committee of Hunan University of Arts and Science (HUAS-20250504). All procedures were performed in accordance with the ethical standards of the responsible institutional committee and with the principles of the Declaration of Helsinki. Informed consent was obtained from all participants prior to their inclusion in the study. Participants were fully informed about the purpose of the study, the procedures involved, and their right to withdraw at any time without penalty. The privacy and confidentiality of all participants were strictly maintained throughout the study. All data were anonymized prior to analysis, and no personally identifiable information was collected, stored, or reported. No financial or material compensation was provided to participants for their participation in this study.

This study successfully designed and validated a novel AI-driven framework for comprehensive pediatric health risk stratification. The primary findings were threefold. First, the framework demonstrates the capability to effectively integrate heterogeneous, multimodal data sources to create a holistic view of a child’s health status. Second, our proposed predictive model, which combines a fine-tuned BERT architecture with an ensemble learning strategy, significantly outperformed established baseline models, such as logistic regression and SVM, in predicting key health risks, achieving an AUC-ROC of 0.85 for early childhood obesity. Third, the prototype system’s risk assessments showed substantial agreement with manual expert evaluations (78% accuracy), confirming its potential clinical utility and feasibility. The successful implementation of the prototype system, featuring intuitive dashboards for both clinicians and parents, further illustrates the practical applicability of this approach in facilitating early and personalized interventions, thereby contributing a novel and robust technological foundation for proactive pediatric health care.

While the presented framework shows considerable promise, certain limitations and avenues for future research warrant discussion. The current validation, though rigorous, was based on a specific dataset of over 40,000 pediatric participants aged 2‐8 years. Broader validation across more diverse pediatric populations and longitudinal follow-up are essential to ascertain long-term predictive accuracy and generalizability. One limitation of our study is the observed slight dip in model performance for the subgroup of children aged 3‐5 years, which we attribute to the relative scarcity of rich, unstructured text data for this age cohort. To mitigate this issue in future work, we plan to incorporate alternative text sources. A promising approach, as suggested, is the inclusion of open-ended responses from parental questionnaires. These narratives, which capture detailed parental concerns and observations, can be processed by our fine-tuned BERT model to generate rich semantic features, thereby enriching the feature set for younger children and addressing the data sparsity issue. A crucial direction for future work is the implementation of advanced bias mitigation techniques. We plan to further explore methods, such as adversarial debiasing, which involves training the model to make predictions that are invariant to sensitive attributes, thereby proactively enhancing the fairness and equity of our risk stratification framework.