Ginder ML remote monitoring

Predicting Malignant Ventricular Arrhythmias Using Real-Time Remote Monitoring

Patient / Population Intervention / Exposure Comparison Outcome

In 2,413 patients with heart failure and an ICD or CRT-D in the IMPACT trial, contributing 59,807 daily remote-monitoring transmissions, a neural-network model of the 30 days of remote-monitoring data before device therapy improved prediction of appropriate ICD therapy for VT/VF compared with multivariable logistic regression.

N
2,413
Design
Post hoc prediction-model analysis of the IMPACT multicentre RCT (patients not randomised to the model)
Endpoint
Prediction of appropriate device therapy for ventricular tachycardia or fibrillation (sensitivity, specificity, AUC)
Relevance
3
ResultNeural 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.
Ginder C, et al. J Am Coll Cardiol. 2023;81(10):949-961. 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.