Risk of heart failure hospitalisation stratified by multisensor alert severity by implantable cardiac devices
L Witten Davodian, D Davodian, F H Fussing, N D Johansen, D Modin, T Bekfani, P Kumarathurai, N Risum, C Chien, M Fudim, A Hernandez, B Claggett, L Koeber, C M Stolen, T Biering-SoerensenAbstract
Background
Remote monitoring alerts based on multisensor data from implanted cardiac devices vary in magnitude and sensor-level composition. Whether within-alert dynamics carry prognostic value beyond binary alert status is unknown.
Purpose
To evaluate whether peak alert severity and individual sensor trajectories during alert episodes stratify risk of 60-day heart failure (HF) hospitalisation.
Methods
In a retrospective analysis, we studied 3,530 alert episodes in 938 patients implanted with a specific multisensor algorithm compatible ICD or CRT-D. The algorithm daily generates a composite HF index based on multiple sensor inputs. When exceeding a fixed, prespecified threshold (16), an alert is triggered and once again resolved when the index recovers below another fixed, prespecified threshold (6). Alert states were defined as continuous periods from onset to resolution. For each alert episode, we identified the peak composite index which was then stratified into peak-index tertiles (T1 <19, T2 19-26, T3 >26). Sensor changes were defined as the difference between the 14-day pre-alert mean and the peak/trough value in five prespecified sensors; respiration rate, thoracic impedance, night heart rate, and first [S1] and third [S3] heart sounds. Sixty-day follow-up was chosen consistent with prior validation studies. The endpoint was acute HF hospitalisation within 60 days from alert onset; death was treated as a competing event. We used cause-specific Cox regression with patient-level clustering, which models the risk of HF hospitalisation while accounting for the competing possibility of death adjusted for age, sex, and device type. Model discrimination was assessed by C-index correction for overfitting via 200 patient-cluster bootstrap resamples.
Results
Median peak index was 22 (IQR 18-29). Sixty-day cumulative incidence increased across peak-index tertiles (T1 <19: 0.9%; T2 19-26: 2.5%; T3 >26: 5.5% (Figure 1B)), with marked risk separation persisting after adjusting for co-variates (T3 vs T1 adjusted HR 5.87, 95% CI 3.16–10.90, p<0.001). T3 alerts accounted for 59% of all HF hospitalisations, whereas fewer than 1% of T1 alerts led to hospitalisation (Figure 1A). All prespecified sensors were independently associated with hospitalisation (HR 1.22–1.62 per SD increase/decrease; all p≤0.004) (Figure 2). Peak index alone yielded C-index 0.706; adding individual sensor changes features did not improve discrimination after overfitting for correction (C-index 0.704).
Conclusions
Peak alert severity indicates varying risk levels beyond binary alert status, while sensor changes map distinct signs of decompensation. These findings support severity-based interpretation and prioritisation of alerts.Multisensor Figure 1Multisensor Figure 2