Short horizon prediction of incident significant arrhythmia from intraoperative heart rate variability, a secondary analysis of the VitalDB Arrhythmia Database

Rhythm disturbance is the most frequent cardiovascular abnormality of the operative period, and new onset atrial fibrillation carries lasting stroke and mortality risk. We tested whether the autonomic profile of a short intraoperative observation window predicts significant arrhythmia in the minutes that follow. Secondary analysis of the VitalDB Arrhythmia Database, with anesthesiologist adjudicated annotations from 482 surgical cases. A 300 s index window was taken at the start of each annotated segment, with the following 300 s as the outcome window. Cases free of significant arrhythmia during the index window were eligible. The endpoint was incident significant arrhythmia, a composite of atrial and ventricular tachyarrhythmias, atrioventricular block and sinus node dysfunction. Logistic regression, random forest and gradient boosting were compared by repeated stratified cross validation. Of 482 cases, 299 were eligible and 109 (36.5%) reached the endpoint. Logistic regression and the random forest discriminated similarly, with areas under the curve of 0.670 (95% CI 0.604 to 0.732) and 0.682 (0.620 to 0.742) (DeLong p = 0.55). The random forest had a favorable calibration slope (0.94), with overprediction of absolute risk, and a negative predictive value of 80.9%. Discrimination was highest for bradyarrhythmic events (0.749). Lower pNN50 and RMSSD, and unexpectedly lower ventricular ectopy, were associated with the endpoint. Ventricular tachyarrhythmia episodes were short, with a median of 4 beats. A five minute heart rate variability profile, derived only from R peak timing and beat labels, predicts significant arrhythmia in the following five minutes of surgery. The signal is physiologically coherent, strongest for bradyarrhythmias, and captured as well by logistic regression as by machine learning. This first prediction benchmark on the VitalDB Arrhythmia Database provides a reproducible foundation for externally validated, clinically enriched intraoperative rhythm prediction.