JMIR Publications

ABSTRACT

Background:

Early and accurate pediatric health risk stratification is critical for optimizing long-term well-being and informing targeted public health interventions. Traditional assessment paradigms often lack comprehensiveness, personalization, and dynamic capabilities. The proliferation of digital health technologies—including electronic health records (EHRs), wearable sensors, and mobile health applications—has generated vast multimodal pediatric datasets. Concurrently, advances in artificial intelligence (AI), particularly natural language processing (e.g., BERT variants) and ensemble learning, offer unprecedented opportunities to analyze complex health interactions. Prior research demonstrates AI's efficacy in predicting conditions like childhood obesity and developmental disorders, yet gaps persist in integrated frameworks leveraging multi-source data for holistic risk assessment. Ethical imperatives surrounding pediatric data privacy (GDPR/HIPAA compliance) and the need for clinical interpretability further underscore the complexity of this domain. This study addresses these challenges by synthesizing EHRs, parental inputs, and sensor data within a unified AI architecture to enable proactive, individualized risk stratification.

Objective:

This study aims to design, implement, and validate an AI-driven framework for comprehensive pediatric health risk stratification. The primary objectives are: (1) To systematically integrate heterogeneous data sources—EHRs, parental questionnaires, wearable device metrics, and environmental factors—into a unified analytical pipeline; (2) To develop and fine-tune domain-specific AI models, including BERT for pediatric knowledge extraction and ensemble methods for risk prediction; (3) To establish clinically actionable risk stratification thresholds across key health domains (e.g., nutrition, mental health, chronic disease); (4) To validate model performance against established clinical benchmarks and baseline algorithms using robust metrics (AUC-ROC, accuracy); and (5) To implement a prototype system with dual-dashboard interfaces for clinicians and caregivers. The framework targets children aged 2–8 years, focusing on prevalent risks such as early childhood obesity, developmental delays, and preventive care adherence. Ultimately, this research seeks to transform fragmented pediatric data into interpretable risk profiles, enabling timely interventions and optimized resource allocation in healthcare systems.

Methods:

The framework employs a multi-tier architecture integrating data acquisition, preprocessing, AI modeling, and risk visualization. Data Integration: Heterogeneous sources (EHRs, wearables, questionnaires) were harmonized using schema mapping and FHIR standards. Preprocessing: Missing values were addressed via MICE imputation; outliers were treated using IQR-based winsorization. Growth parameters were normalized to age/sex Z-scores (Equation 2). Feature Engineering: Domain-specific features included nutritional intake indices, developmental velocity metrics, and cumulative environmental exposures. SHAP values guided feature selection. AI Models: A BERT model fine-tuned on pediatric medical literature extracted entities from clinical notes. An ensemble of Gradient Boosting and Random Forest classifiers predicted risks, optimized via 5-fold cross-validation. Risk Stratification: Continuous risk scores were discretized into four tiers (Low to Very High) using percentile-based thresholds validated by pediatric experts. System Design: A Python-based backend (TensorFlow, Scikit-learn) hosted microservices; PostgreSQL/MongoDB managed data storage; React.js powered clinician/parent dashboards. Validation: Models were trained on 28,000 subjects (70%), validated on 6,000 (15%), and tested on 6,000 (15%). Performance was benchmarked against logistic regression, SVM, and single Decision Trees using AUC-ROC, precision-recall curves, and F1-scores.

Results:

The AI framework demonstrated significant efficacy in risk stratification across 40,000 pediatric subjects (aged 2–8 years). For early obesity prediction, the ensemble model achieved an AUC-ROC of 0.85 (95% CI: 0.82–0.88), surpassing logistic regression (AUC-ROC=0.72) and SVM (AUC-ROC=0.75) (Table 4). Precision-recall analysis yielded an AUC-PR of 0.70. Model sensitivity and specificity were 0.78 and 0.80, respectively, with a Brier score of 0.15 indicating high calibration. Automated risk assessments aligned with manual expert evaluations at 78% accuracy (Figure 6). Subgroup analyses revealed consistent performance across age strata (<2 years: AUC-ROC=0.84; 2–5 years: 0.85; >5 years: 0.86) and socioeconomic levels (Low SES: AUC-ROC=0.83; High SES: 0.86) (Table 5). The BERT-based knowledge extractor identified pediatric disease entities with an F1-score of 0.89. SHAP analysis highlighted key predictors: BMI Z-scores (mean |SHAP|=0.32), sedentary duration (0.28), and sugary beverage intake (0.25). The prototype system processed multi-modal data with a mean latency of 8.2 seconds per patient, enabling real-time dashboard updates for clinicians and parents. Comparative analysis confirmed the ensemble's superiority over all baselines (DeLong’s test, p<0.05).

Conclusions:

This study establishes a validated AI framework for comprehensive pediatric health risk stratification by integrating multi-modal data and advanced machine learning. Key contributions include: (1) A scalable architecture harmonizing EHRs, wearable metrics, and parental inputs; (2) A fine-tuned BERT model enabling pediatric knowledge extraction from unstructured texts; (3) An ensemble predictor demonstrating robust performance (AUC-ROC=0.85 for obesity risk); and (4) Clinically interpretable risk tiers facilitating targeted interventions. The prototype’s dual-dashboard system successfully translates complex data into actionable insights for stakeholders. Despite strengths, limitations include dataset age restrictions (2–8 years) and the need for longitudinal validation. Future work should: (1) Extend coverage to adolescence and infancy; (2) Incorporate real-time data streams for dynamic risk updating; (3) Enhance model interpretability via granular explainable AI (XAI) techniques; and (4) Evaluate clinical impact through prospective trials measuring intervention efficacy. This framework provides a foundation for data-driven, personalized pediatric care, potentially reducing long-term health burdens through early risk mitigation. Clinical Trial: This study utilized retrospective data analysis and did not constitute a clinical trial involving prospective interventions; thus, formal trial registration was not applicable. The dataset comprised de-identified records from 40,000 pediatric subjects (aged 2–8 years), sourced from electronic health records, parental surveys, and wearable devices between 2018–2022. Data acquisition complied with institutional review board protocols at Hunan University of Arts and Science (approval #HUAS-IRB-2023PED01) and adhered to GDPR/HIPAA standards for pediatric data anonymization. Parental consent was obtained for questionnaire and wearable data collection via opt-in digital agreements. External validation employed an independent test set (n=6,000) with rigorous separation from training/validation phases to prevent leakage. Model performance was benchmarked against established clinical assessments through blinded expert reviews (n=100 case pairs), achieving 78% concordance in risk categorization. All code and de-identified metadata are archived in a Zenodo repository (DOI: 10.5281/zenodo.XXXXXX) for reproducibility. Future prospective trials evaluating clinical implementation will be registered prospectively in the Chinese Clinical Trial Registry (ChiCTR), with protocols addressing intervention efficacy, cost-effectiveness, and long-term health outcome monitoring.