Ginder ML remote monitoring
2023
Predicting Malignant Ventricular Arrhythmias Using Real-Time Remote Monitoring
In 2,413 patients with heart failure and an ICD or CRT-D, a neural network reading the 30 days of remote-monitoring data before device therapy improved prediction of appropriate ICD therapy for VT/VF compared with multivariable logistic regression.
Result
Primary endpoint: Prediction of appropriate device therapy for ventricular tachycardia or fibrillation (sensitivity, specificity, AUC)
Neural network AUC 0.90 (sensitivity 54%, specificity 96%) vs logistic regression AUC 0.72 (sensitivity 39%, specificity 91%); P<0.01 for the comparison. Appropriate therapies (141 shocks, 10 antitachycardia pacing) occurred in 151 patients; IMPACT had randomised 2,718.
- Subtopic
- Rhythm & devices
- Design
- Post hoc prediction-model analysis of the IMPACT multicentre RCT (patients not randomised to the model)
- Relevance
- Supporting: context rather than must-know.
- Source
- Ginder C, et al. J Am Coll Cardiol. 2023;81(10):949-961.doi 10.1016/j.jacc.2022.12.024
Discussion & critique
Shows that the daily data ICDs already transmit, lead impedance, ectopy, heart rate, activity, carry a predictive signal for malignant arrhythmia a month ahead, and that a neural network extracts it far better than regression. PubMed indexes it as an RCT because the data come from IMPACT, but patients were not randomised to any model: it is a post hoc, retrospective prediction analysis without external validation or a prospective intervention, so it is hypothesis-generating.