Individualized baseline modelling of implantable cardioverter defibrillators remote monitoring data detects deviations preceding ventricular arrhythmias
T Ito, T Noda, M Nakano, N Ueda, H Sato, S Kumabe, S Yamanaka, N Yamamoto, T Chiba, K Nochioka, S Kato, T Asakura, K Kusano, T Miyata, S YasudaAbstract
Background
Ventricular tachycardia (VT) and ventricular fibrillation (VF) remain major causes of morbidity in patients with implantable cardioverter defibrillators (ICDs), often occurring without overt preceding clinical deterioration. Contemporary ICD systems continuously collect multidimensional physiological data through remote monitoring, offering an opportunity for early detection of arrhythmic risk. However, prediction of ventricular arrhythmias remains challenging because heterogeneity among individuals limits the identification of generalizable risk markers, and single physiological parameters have shown limited predictive value. In contrast, physiological status within individual patients is relatively stable over short periods, suggesting that deviations from baseline conditions for each patient rather than absolute values may precede ventricular arrhythmias. We therefore hypothesized that departures from individualized baseline patterns could signal impending VT/VF events.
Purpose
To evaluate whether deviations from stable individualized conditions derived from daily ICD remote monitoring data are temporally associated with VT/VF.
Methods
This multicenter retrospective study analyzed daily remote monitoring data from 46 patients with an ICD or cardiac resynchronization therapy defibrillator, including 14 with appropriate ICD therapy and 32 without. A convolutional neural network autoencoder was trained using physiological parameters from patients without ICD therapy to learn stable baseline patterns based on consecutive 5-day windows from 90-day recordings (Figure1). The model was applied to two separate 90-day periods in patients with ICD therapy: (1) preceding VT/VF and (2) a temporally distinct control period. Deviations were quantified as mean absolute reconstruction error (MAE). Daily MAE values were generated using sliding 5-day windows shifted by one day. The threshold for significant deviation was defined as the 95th percentile of MAE obtained from patients without ICD therapy using patient-level cross-validation.
Results
Physiological deviations increased preferentially during periods preceding ventricular arrhythmias. Median MAE rose from 0.53 to 1.45 prior to VT/VF, whereas only a modest change was observed during control periods (0.50 to 0.74) (Figure 2A). Five consecutive threshold exceedances provided the best discrimination between periods prior to VT/VF and control periods (sensitivity 0.64, specificity 0.86, positive predictive value 0.82, negative predictive value 0.71). Using this criterion, 9 of 14 patients demonstrated significant deviations before VT/VF, with a median interval of 42 days from the first exceedance to VT/VF onset (Figure 2B).
Conclusion
Deviations from learned individualized baseline physiological patterns were temporally associated with subsequent VT/VF events. Daily ICD remote monitoring data may enable personalized risk surveillance and warrants prospective validation.