Multimodal EHR-Based Prediction of Pediatric Asthma Exacerbations

This study leverages UF Health electronic health records to demonstrate that an interpretable XGBoost model, integrating clinical notes and medication data, effectively predicts pediatric asthma exacerbations over 6-, 12-, and 24-month horizons, offering a promising framework for risk stratification and clinical decision support.

The Gist

Imagine a child's asthma as a storm cloud hanging over their head. Sometimes the cloud is small and harmless, but other times, it bursts into a heavy rainstorm (an "exacerbation") that sends the child rushing to the Emergency Room or the hospital.

For a long time, doctors have tried to predict when these storms will hit, but their tools have been a bit like looking at a map with only half the terrain drawn. They could see the "structured" data—like the child's age, their diagnosis codes, and what medicines they bought—but they were missing the most important part of the story: what the doctors actually wrote in their notes.

This paper is about building a super-weather forecast for pediatric asthma by combining the map with the story.

The Problem: The "Blind Spot" in Medical Records

Think of a child's medical record as a giant library.

• The Structured Data: These are the library's index cards. They tell you the book's title (diagnosis), the author (medication), and the date it was checked out. Computers are great at reading these cards.
• The Clinical Notes: These are the actual pages of the book. Doctors write things like, "The child is wheezing heavily at night," or "They used their rescue inhaler three times today." These details are crucial, but they are buried in free-flowing text that traditional computers often ignore.

Previous studies tried to predict asthma attacks using only the index cards. This paper argues that to get a real forecast, you need to read the pages too.

The Solution: A "Multimodal" Detective Team

The researchers at the University of Florida built a digital detective team using Machine Learning (AI). They didn't just look at one type of clue; they looked at everything.

1. The Data Source: They gathered the medical records of over 56,000 children (ages 2–18) from 2011 to 2023. That's like reading the medical history of a small city's entire youth population.
2. The Two "Detectives" (Phenotypes): They used two different methods to identify which children actually had asthma:
   • CAPriCORN: The "Strict Accountant." It only counts children if their records have specific diagnosis codes and medication lists.
   • COMPAC: The "Storyteller." It looks at the codes and scans the doctors' notes for keywords like "wheeze," "cough," or "shortness of breath."
3. The Prediction Windows: They tried to predict storms for three different timeframes: 6 months, 1 year, and 2 years into the future.

The Secret Weapon: Reading the Notes

The researchers used a special AI tool (called GatorTron) that acts like a super-fast librarian. It scanned thousands of doctors' notes to find specific phrases about breathing problems.

They then fed this information into a "brain" (an AI model called XGBoost). Think of XGBoost as a very smart, experienced meteorologist who is better at spotting patterns than any other model they tried (like Logistic Regression or Random Forest).

What Did They Find?

The results were like finding a crystal ball that actually works.

• The Best Model: The XGBoost model was the clear winner. It could predict an asthma attack with about 80–85% accuracy (AUC scores), which is very high for medical predictions.
• The Most Important Clues: When the researchers asked the AI, "What made you think this storm was coming?" the AI pointed to three main things:
  1. Symptoms in the Notes: Words like "wheeze," "cough," and "trouble breathing" found in the doctors' handwritten notes were the strongest predictors. This proved that reading the notes is essential.
  2. Rescue Meds: If a child was using their "rescue inhaler" (albuterol) frequently, the risk of a big storm went up.
  3. Allergies: Children with allergies (like hay fever) were more likely to have asthma attacks.

Interestingly, the AI also noticed that children taking ibuprofen seemed to have higher risk (though this might just mean they were already sick with a virus), and those taking cetirizine (an allergy pill) seemed safer, likely because their allergies were being managed.

Why Does This Matter?

Imagine if your child's doctor could get a text message saying, "Warning: Based on your child's recent notes and medication use, there is a high chance of an asthma attack in the next 6 months."

This isn't about replacing doctors; it's about giving them a super-powered early warning system.

• Proactive Care: Instead of waiting for the child to rush to the ER, the doctor could call the family, adjust their medication, or reinforce their "Asthma Action Plan."
• Fewer Trips: This could mean fewer emergency room visits, fewer missed school days, and less stress for families.

The Catch (Limitations)

The authors are honest about the limitations. They only looked at data from one hospital system (University of Florida). It's like testing a weather forecast model only in Florida; it might not work perfectly in Alaska yet. They also need to make sure the AI doesn't accidentally treat different groups of people unfairly.

The Bottom Line

This paper shows that to predict a child's asthma storm, you can't just look at the checklist; you have to read the story. By teaching computers to understand both the medical codes and the doctors' notes, we can build a smarter, more accurate safety net for children with asthma, helping them breathe easier and stay out of the hospital.

Technical Summary

1. Problem Statement

Pediatric asthma exacerbations are a leading cause of emergency department (ED) visits and hospitalizations, imposing significant economic and health burdens. Despite this, accurate risk prediction remains challenging due to:

• Lack of Consensus: No widely accepted clinical prediction tools exist for children.
• Data Limitations: Existing models often rely on adult cohorts or mixed-age groups, failing to account for pediatric-specific pathophysiology and triggers.
• Modality Gaps: Most studies utilize only structured Electronic Health Record (EHR) data (diagnoses, vitals, labs) or perform "shallow" integration of unstructured text. They often miss critical symptom details (e.g., nocturnal cough, rescue inhaler overuse) captured only in clinical notes.
• Generalizability Issues: Many models suffer from site-specific coding variability, small sample sizes, and sparse outcome labels, limiting their applicability across diverse populations.

2. Methodology

Data Source and Cohort

• Data: Longitudinal EHR data from the University of Florida (UF) Health system (2011–2023), covering over 2 million patients.
• Population: Children aged 2–18 years. Children under 2 were excluded to avoid confounding from bronchiolitis.
• Phenotyping: Two computable phenotypes (CPs) were applied to identify asthma patients:
  1. CAPriCORN: A rule-based algorithm using structured data (diagnosis codes + medication exposure).
  2. COMPAC: A hybrid approach combining structured criteria with text-derived evidence (asthma signs/symptoms extracted via keyword rules from clinical notes).
• Outcome Definition: An asthma exacerbation was defined as a composite endpoint meeting any of the following:
  • ED visit with an asthma diagnosis.
  • Hospital admission with an asthma diagnosis.
  • Outpatient visit with an asthma diagnosis plus a systemic corticosteroid prescription linked to the encounter.

Feature Engineering (Multimodal Integration)

The study integrated three distinct data modalities:

1. Structured Diagnoses: ICD-9/10 codes mapped to Phecodes to reduce dimensionality.
2. Medications: Aggregated at the ingredient (generic) level.
3. Clinical Notes (NLP):
   • Utilized GatorTron (a clinical BERT model) for Named Entity Recognition (NER) restricted to problem entities.
   • Extracted asthma-related symptom mentions (e.g., wheeze, cough, dyspnea).
   • Terms were normalized, ranked by frequency, and refined by clinician review to create a set of 1,000 keywords.
   • Term frequencies served as input features.

• Filtering: Features with <1% prevalence in the cohort were excluded to reduce sparsity.

Prediction Windows and Models

• Timeline: The index date was the first asthma diagnosis encounter. Prediction windows were set at 6, 12, and 24 months post-index.
• Algorithms: Five supervised ML models were evaluated:
  • Lasso Regression (LASSO)
  • Logistic Regression (LR)
  • Multilayer Perceptron (MLP)
  • Random Forest (RF)
  • XGBoost (Extreme Gradient Boosting)
• Training Strategy: Stratified five-fold cross-validation with Bayesian hyperparameter optimization. Thresholds were optimized using Youden's J statistic.
• Evaluation Metrics: Primary metric was Area Under the Receiver Operating Characteristic Curve (AUC). Secondary metrics included Accuracy, F1-score, Precision, Recall, and Specificity.
• Interpretability: SHapley Additive exPlanations (SHAP) were used to quantify feature importance and directionality.

3. Key Results

Cohort Characteristics

• Sample Size: ~27,800 (CAPriCORN) and ~28,800 (COMPAC) children.
• Demographics: ~57% Male, ~43% Female; Non-Hispanic Black (43%) and Non-Hispanic White (38%) were the largest groups.
• Event Rates: Approximately 10% of patients experienced an exacerbation within 6 months, rising to ~20% within 24 months.

Model Performance

XGBoost consistently outperformed all other models across both phenotypes and all time horizons.

• 6-Month Horizon:
  • COMPAC: AUC 0.813 (XGBoost) vs. 0.795 (MLP).
  • CAPriCORN: AUC 0.827 (XGBoost) vs. 0.810 (MLP).
• 12-Month Horizon:
  • COMPAC: AUC 0.844 (XGBoost).
  • CAPriCORN: AUC 0.813 (XGBoost).
• 24-Month Horizon:
  • COMPAC: AUC 0.829 (XGBoost).
  • CAPriCORN: AUC 0.848 (XGBoost).

Feature Importance (SHAP Analysis)

The most predictive features were consistently derived from unstructured clinical notes and rescue medication usage:

1. Symptom Terms: Keywords indicating dyspnea on exertion, wheeze, cough, and asthma.
2. Medications: Use of rescue medication (albuterol sulfate) was a strong risk indicator. Conversely, antihistamine use (cetirizine) was associated with lower risk.
3. Physical Exam/Comorbidities: Accessory muscle use, nasal flaring, and allergy diagnoses.

• Stability: The ranking of these predictors remained stable across the 6, 12, and 24-month windows.

4. Key Contributions

1. Multimodal Framework: Demonstrated that integrating NLP-extracted symptom trajectories from clinical notes with structured EHR data significantly improves prediction accuracy over structured-only models.
2. Pediatric Specificity: Developed and validated models specifically for the 2–18 age group, addressing the gap in pediatric-focused asthma prediction tools.
3. Phenotype Comparison: Systematically compared two distinct computable phenotypes (CAPriCORN vs. COMPAC), showing that text-integrated phenotyping (COMPAC) captures risk factors effectively, though both yielded similar model performance trends.
4. Interpretability: Provided a transparent view of risk drivers using SHAP, confirming that clinical intuition (symptoms + rescue meds) aligns with model predictions.

5. Significance and Future Directions

• Clinical Impact: The model supports individualized, multi-horizon risk stratification. Clinicians could use these scores to trigger proactive outreach, step-up therapy, or reinforce asthma action plans for high-risk children, potentially reducing ED visits and hospitalizations.
• Scalability: The framework is designed to be embedded within EHR systems for real-time decision support.
• Limitations & Next Steps:
  • The study was retrospective and single-site; external validation across different health systems is required.
  • Future work must address fairness and calibration across subgroups (race, sex, insurance).
  • Prospective implementation studies (e.g., silent-mode testing followed by intervention trials) are needed to prove clinical utility and cost-effectiveness.
  • Further investigation is needed into specific medication associations (e.g., ibuprofen) to distinguish between confounding by indication and true causal effects.

In conclusion, this study establishes a robust, interpretable, text-integrated machine learning framework for predicting pediatric asthma exacerbations, offering a pathway toward more precise, preventative asthma management.