Dynamic Prediction of Radiation-Induced Arrhythmia Using Landmark Modeling with Longitudinal Heart Rate and Cardiac Substructure Dosimetry
Abstract
Purpose
Radiation therapy–associated cardiac arrhythmia is a clinically significant complication of thoracic RT, with important consequences for long-term cardiovascular health and treatment outcomes. Although arrhythmia risk evolves over time and is linked to cardiac radiation exposure, no established framework currently exists for longitudinal risk monitoring or adaptive RT planning. We propose a landmark-based modeling approach that integrates longitudinal pulse dynamics with time-scaled cardiac substructure dosimetry to enable dynamic risk stratification and evaluate its potential for cardioprotective mid-treatment intervention.
Methods
We analyzed 112 NSCLC patients receiving thoracic chemoradiation followed by immunotherapy (Durvalumab), of whom 21 (19%) developed (de novo) post-RT arrhythmia. We retrieved longitudinal pulse rates data and AI-assisted cardiac substructure dose and applied landmark modeling to dynamically predict arrhythmia risk at five intermediate landmarks within RT course (10-day intervals spanning days 10-50), incorporating longitudinal pulse features (7-day rolling mean, maximum, variability, cumulative excess) and time-scaled cardiac substructure doses. Repeated stratified cross-validation compared the multimodality models. Baseline clinical features (age, sex, beta-blocker use, hypertension) defined risk-stratified thresholds for adaptive simulation.
Results
Pulse+Dose model achieved peak AUC of 0.82±0.18 at day 20, suggesting early dynamic signal, whereas Clinical+Dose model provided stable prediction (AUC 0.76–0.79) across all landmarks. Stable dosimetric predictors included max- and volumetric doses to the whole heart, conduction system (sinoatrial and atrioventricular nodes), and ventricles. In adaptive simulation using clinical-stratified thresholds (0.60–0.80), 67% (14/21) of eventual arrhythmia cases were flagged during RT, with 79% (50/63) of triggers happening within the first 20 days when dose modification remains feasible. Simulated cardiac dose adjustment demonstrated arrhythmia risk reduction in all flagged cases (median: 39.5%, range: 3.4%-81%).
Conclusion
A framework combining pulse dynamics and cardiac dosimetry enables early identification of elevated arrhythmia risk during NSCLC chemoradiation. Concentration of triggers in early (~20 fractions) suggest a clinically actionable window for adaptive replanning.