Predicting Infant Nonattendance at the Next Recommended Well-Child Visit: Model Development and Validation

Background: Well-child visits (WCVs) are essential for monitoring infant development, yet nonattendance remains common. Machine learning models have predicted appointment no-shows, but these approaches require a visit to be scheduled and cannot identify children whose visits are never scheduled or are cancelled in advance. We developed and validated models that predict infant nonattendance at the next recommended WCV using information available at the current visit. Methods: We used visit-level electronic health record data from two demographically distinct pediatric practices: Practice A (5,300 visits, 1,215 patients) served as the training and internal validation set; Practice B (5,274 visits, 1,271 patients) served as the external validation set. We compared regularized logistic regression, random forest, and XGBoost using group k-fold cross-validation. Model discrimination was assessed using AUC and average precision, and threshold-dependent metrics were reported at the F1-maximizing threshold. We evaluated subgroup performance by race/ethnicity, preferred language, and insurance type in the external validation set. Results: All three models achieved similar discrimination in internal validation (AUC 0.69-0.72, average precision 0.36-0.41) and external validation (AUC 0.66-0.68, average precision 0.36-0.38). Across all models, care delivery features, including visit timepoint, visit delay, scheduling lead time, and prior no-shows, were the most influential predictors. Logistic regression with LASSO regularization retained six nonzero coefficients, suggesting comparable performance is achievable with a reduced feature set. Model discrimination was similar across insurance groups, between English- and Spanish-speaking patients, and between Hispanic and non-Hispanic Black patients relative to non-Hispanic white patients. Estimates for smaller subgroups were imprecise due to limited sample sizes. Conclusions: Predicting WCV nonattendance at the next recommended timepoint is feasible using routine visit data. A parsimonious logistic regression model performed comparably to more complex algorithms. Because prediction occurs during an active visit, clinicians can address barriers to the next visit while the family is still present.