Perinatal & Pediatric
Predicting Infant Nonattendance at the Next Recommended Well-Child Visit: Model Development and External Validation Amanda Luff* Amanda Luff Luff Luff Luff Luff Luff Advocate Health
Background: Well-child visits (WCVs) are critical for infant preventive care, yet missed visits are common and are associated with delayed diagnoses and increased acute care use. A tool to predict next-visit nonattendance at the current visit could enable proactive support.
Methods: Using EHR data from two Chicago-area pediatric clinic groups, we developed models to predict nonattendance at the next recommended WCV among infants who attended the current WCV. Clinic A (training/internal validation) contributed 1,215 patients (3,654 visits) and Clinic B (external validation) 1,271 patients (3,044 visits). Features captured visit context (timepoint, visit delay, scheduling lead days, provider type, new to clinic), prior utilization (ED use, prior no-show, prior missed WCV, immunization refusal), and patient indicators (sex, insurance, in-system birth; low birth weight, preterm birth, NICU admission). We compared regularized logistic regression, random forest, and XGBoost. Nested group k-fold cross-validation (by patient) optimized average precision. We estimated SHAP values and report threshold-specific metrics using F1-maximizing cutoffs.
Results: Missed-next-visit rates were 16.2% (Clinic A) and 20.7% (Clinic B). Discrimination and average precision were modest and similar across models: internal AUC 0.69–0.72 and average precision 0.36–0.41; external AUC 0.66–0.68 and average precision 0.36–0.38. At the F1-maximizing threshold, recall (sensitivity) ranged 0.40–0.51 internally and 0.54–0.71 externally, with precision (positive predictive value) 0.30–0.37; negative predictive value remained high (~0.86–0.89). In all models, SHAP values identified timepoint, visit delay, schedule lead days, and prior no-show as among the top 6 most influential features.
Conclusion: Predicting next-WCV attendance is feasible using routinely available EHR features, though performance is modest. A parsimonious LASSO logistic regression model performed comparably to more complex models.