Nationwide EHR-Based Chronic Rhinosinusitis Prediction Using Demographic-Stratified Models

Technical Summary: Nationwide EHR-Based Chronic Rhinosinusitis Prediction Using Demographic-Stratified Models

Problem Statement
Chronic rhinosinusitis (CRS) is a heterogeneous inflammatory disorder affecting approximately 11–12% of the U.S. population, imposing significant morbidity and financial burdens. Early identification is challenging because symptoms overlap with common conditions like allergic rhinitis, and formal diagnosis typically requires objective evidence from nasal endoscopy or CT scans, which are not always available in primary care settings. Prior predictive studies have relied on single-institution cohorts, limiting population-level generalizability, or utilized resource-intensive modalities like genetic sequencing and imaging. Furthermore, existing machine learning approaches often fail to account for the pronounced demographic heterogeneity in CRS phenotypes, such as age-dependent shifts in pathology and sex-specific differences in disease presentation.

Methodology
The authors developed a predictive framework using nationwide longitudinal Electronic Health Record (EHR) data from the NIH "All of Us" Research Program, standardized to the OMOP Common Data Model. The methodology consists of three core components:

1. Data Preprocessing and Cohort Construction:

   • Target Definition: CRS cases were identified using a rule-based phenotyping strategy requiring at least two qualifying diagnostic codes within a two-year period. Controls were propensity-score matched to cases based on age, sex, race, ethnicity, and visit frequency to mitigate surveillance bias.
   • Feature Engineering: Raw transactional logs were transformed into a recency-based feature representation. For each clinical variable (conditions, medications, procedures), the feature value represents the number of days between the most recent event and the index date. Events outside the two-year window or never occurring were assigned a sentinel value (999,999) to allow tree-based models to distinguish missing data without imputation.
   • Balancing: A 1:1 nearest-neighbor matching strategy was employed to create a balanced cohort of 17,560 participants (8,780 cases and 8,780 controls).

2. Hybrid Feature Selection:
   To address extreme sparsity and dimensionality (over 110,000 candidate codes), a two-stage pipeline was implemented:

   • Stage 1 (Statistical Coarse Screening): Variables were ranked by the absolute difference in prevalence between cases and controls ($|P_{target} - P_{control}|$). The top 1,100 features (300 conditions, 500 procedures, 300 medications) were retained.
   • Stage 2 (Model-Based Fine Screening): An optimized global XGBoost classifier was trained on the 1,100 features. Features were ranked by information-gain-based importance, and the top 100 were selected. This subset accounted for 80.6% of the total cumulative SHAP value, ensuring the preservation of predictive power while enhancing interpretability.

3. Demographic-Stratified Modeling:
   Recognizing that CRS phenotypes vary by sex and age, the cohort was partitioned into six mutually exclusive strata: Male/Female crossed with three age groups (18–40, 40–60, 60+). Six independent XGBoost models were trained, one for each stratum. Each model underwent independent hyperparameter optimization using the Bayesian Tree-structured Parzen Estimator (TPE) algorithm over 100 trials to maximize AUC-ROC.

Key Contributions

• Nationwide Scale: Leveraged the diversity and richness of the "All of Us" cohort to overcome the generalizability limitations of single-institution studies.
• Hybrid Feature Selection: Introduced a pipeline that compresses >110,000 clinical codes into 100 interpretable features, balancing parsimony with predictive performance.
• Demographic Stratification: Proposed a modeling framework that explicitly accounts for age-dependent sexual dimorphism in disease phenotypes, training six tailored models rather than a single global model.
• Clinical Applicability: Demonstrated that routinely collected EHR data can support risk stratification without reliance on expensive imaging or genetic data.

Results
The proposed demographic-stratified XGBoost framework achieved an overall Area Under the Curve (AUC) of 0.8461. This represented an improvement of 0.0168 over the best baseline (a global XGBoost model with AUC 0.8293). The proposed method also outperformed Random Forest (0.8159), SVM (0.8120), and Deep Learning (0.8072) models.

Crucially, the stratified approach improved discrimination across all six subgroups, with AUC gains ranging from +0.0076 to +0.0278. The model achieved a sensitivity of 0.6889, significantly surpassing standard machine learning approaches, which is clinically advantageous for screening to ensure affected patients are not overlooked.

Analysis of feature importance revealed distinct biological signals across groups:

• Males: Models prioritized structural and polypoid features (e.g., deviated nasal septum in younger males, nasal polyps in older males).
• Females: Models favored mucosal inflammation and atopy markers (e.g., allergic rhinitis, asthma, and infectious markers in older females).

Significance
The paper claims that explicitly modeling demographic heterogeneity yields measurable gains in CRS risk prediction. By capturing the specific risk factors associated with different sex-age subgroups, the framework offers a more accurate tool for population-representative risk stratification. The authors suggest that this approach can facilitate earlier triage and referral prioritization in primary care settings, particularly when diagnostic imaging is unavailable, shifting decision-making from subjective impression to evidence-driven prediction. The work underscores the utility of routine EHR data in identifying potential CRS cases before specialist evaluation.