Author(s): GONZALEZ, J., NACHER, V., GARCIA, D., MURILLO, I., VIZCAINO, G., JIMENEZ, C., Institution: BARCELONA INSTITUTE FOR GLOBAL HEALTH (ISGLOBAL) - ESG65341695, Country: SPAIN, Abstract-ID: 1304

INTRODUCTION:
Heart rate variability (HRV), is widely used in endurance sports to monitor recovery and guide training prescription. In applied settings, short rolling averages (e.g., 7 days) are typically used to interpret autonomic status, implicitly assuming immediate and linear responses to training load. However, physiological adaptations to accumulated load are delayed and non-linear. Distributed Lag Non-Linear Models (DLNM) allow simultaneous estimation of non-linear exposure–response relationships and their temporal lag structure. This study applied DLNM to model and prospectively predict daily RMSSD responses to training load, and to quantify model-derived autonomic suppression.
METHODS:
Daily RMSSD (outcome) and Training Stress Score (TSS; exposure) were collected in 12 endurance athletes (7 elite, 5 amateur). A DLNM with a maximum lag of 21 days was fitted to capture delayed and cumulative effects of TSS on RMSSD. To determine the monitoring duration required for stable forecasting, time series were progressively truncated (15–240 days). Within each scenario, models were refitted using a 70/30 chronological split (train/test), and the first 7 days of the test segment were predicted out-of-sample. Prediction performance was quantified using mean absolute error (MAE). Once calibrated, the model was used to simulate RMSSD trajectories under hypothetical weekly TSS prescriptions. Autonomic suppression was operationally defined as the model-predicted reduction in RMSSD relative to individual baseline, quantified as the cumulative area below baseline across the simulated microcycle.
RESULTS:
DLNM identified structured lag-response patterns, with peak reductions in RMSSD occurring several days after high TSS exposure, confirming delayed autonomic effects. With ≤60 days of monitoring, no valid 7-day predictions were possible. Prediction accuracy improved markedly beyond 120 days (≈44 ms at 120 days; ≈19 ms at 150 days) and stabilized between 180–240 days (≈15–18 ms).
Simulation analyses revealed that weekly training prescriptions with identical total TSS produced substantially different levels of predicted cumulative autonomic suppression depending on intra-week load distribution. Greater within-week load variability was associated with larger predicted HRV suppression and increased probability of at least one fatigue day, despite equivalent total weekly load.
CONCLUSION:
DLNM provides a physiologically coherent and actionable framework for HRV monitoring. Beyond day-to-day prediction, this approach enables prospective quantification of expected autonomic suppression under planned training scenarios while accounting for delayed and cumulative load effects. Reliable forecasting required at least 3–4 months of monitoring to characterize athlete-specific lag structures. These findings suggest that weekly load distribution—not only total load—modulates predicted autonomic stress, supporting the integration of distributed lag modeling into predictive performance analytics system.