Machine learning-derived cycle length variability metrics predict spontaneously terminating ventricular tachycardia...

Implantable cardioverter defibrillator (ICD) therapies have been associated with increased mortality and should be minimized when safe to do so. We hypothesized that machine learning-derived ventricular tachycardia (VT) cycle length (CL) variability metrics could be used to discriminate between sustained and spontaneously terminating VT.

In this single-centre retrospective study, we analysed data from 69 VT episodes stored on ICDs from 27 patients (36 spontaneously terminating VT, 33 sustained VT). Several VT CL parameters including heart rate variability metrics were calculated. Additionally, a first order auto-regression model was fitted using the first 10 CLs. Using features derived from the first 10 CLs, a random forest classifier was used to predict VT termination. Sustained VT episodes had more stable CLs. Using data from the first 10 CLs only, there was greater CL variability in the spontaneously terminating episodes (mean of standard deviation of first 10 CLs: 20.1 ± 8.9 vs. 11.5 ± 7.8 ms, P < 0.0001). The auto-regression coefficient was significantly greater in spontaneously terminating episodes (mean auto-regression coefficient 0.39 ± 0.32 vs. 0.14 ± 0.39, P < 0.005). A random forest classifier with six features yielded an accuracy of 0.77 (95% confidence interval 0.67 to 0.87) for prediction of VT termination.

Ventricular tachycardia CL variability and instability are associated with spontaneously terminating VT and can be used to predict spontaneous VT termination. Given the harmful effects of unnecessary ICD shocks, this machine learning model could be incorporated into ICD algorithms to defer therapies for episodes of VT that are likely to self-terminate.