Explanatory Versus Predictive Modeling in Heart Rate Variability Research: Nonlinearity, Model Choice, and Machine Learning

Heart rate variability (HRV) changes nonlinearly across the lifespan, and such nonlinearity is often taken to require machine learning (ML). Distinguishing explanatory from predictive modeling, we tested three propositions using 24 h Holter recordings from a clinically referred population (ALLSTAR database), sampled uniformly across ages 0–102 years (n = 3854). Explanatory models expressed 12 HRV indices as functions of age and sex; predictive models estimated age, an error-free benchmark, from HRV under nested, age-stratified cross-validation. First, all indices showed significant nonlinear age dependence, but general linear models with spline terms explained it as well as random forest and gradient boosting (cross-validated R2 differences ≤ 0.02) while providing parameters, confidence intervals, and tests. Second, gradient boosting reduced the mean absolute error of linear regression from 11.3 to 10.1 years (95% CI of the reduction in repeated analyses, 0.8–1.3 years), with gains largest in children and older adults, where all models had their largest errors and bias, and reduced to 0.3 years under the database age distribution. Third, permutation importance agreed only moderately with explanatory age dependence (Spearman ρ 0.48–0.73), differed between models and permutation schemes, and was allocated within correlated index groups in model-specific ways. In this clinically referred, cross-sectional dataset, nonlinearity did not imply that ML was required, and variable importance should not be read as physiological importance.