PO-03-226 MACHINE LEARNING IDENTIFIES FEATURES OF ARRHYTHMOGENIC QRS FRAGMENTATION IN HUMAN CARDIOMYOPATHY: IMPLICATIONS FOR IMPROVING AND AUTOMATING RISK STRATIFICATION

Abstract

Slow, heterogeneous ventricular activation can provide the substrate for reentrant ventricular arrhythmias (VA), but its manifestation on the surface ECG as a risk stratifier is not well-defined beyond the presence of QRS duration (QRSd) >120ms or QRS fragmentation (fQRS). Our aim was to use machine learning (ML) to characterize the spatiotemporal features of QRS fragmentation that best predict VA in patients with cardiomyopathy (CM). Prospectively enrolled ischemic and non-ischemic CM patients with prophylactic defibrillators (n=95) underwent digital, high-resolution 12-lead ECG recordings for 5-minutes during intrinsic rhythm. Abnormal intra QRS peaks in the signal-averaged precordial leads (V1-V6) were identified using our previously validated time-domain algorithm. For each spatial location (V1-V6), QRS peak characteristics including width, amplitude and relative temporal location within the QRS were transformed into 4-bin histograms. Random forest models of the QRS peak counts and characteristics in each lead were trained on groups of 76 patients and tested on groups of 19 patients with stratified repeated random subsampling cross-validation to determine the features that predicted VA outcome. QRS peak counts and characteristics were additionally modelled with QRSd and other relevant clinical characteristics. CM patients (age 62±11 yrs, male 85%, LVEF 27±7%, ischemic etiology 62%) were followed for 24 (15-43) months and 22% had VA (+VA). QRSd was longer in +VA than -VA (135±25 vs. 114±26ms, p=0.001), but age, LVEF, CM etiology and fQRS were similar. The individual ML models determined (ROCAUC) QRS peak count (0.87), width (0.88) and location (0.84) to be predictors of VA (Table). In these models, the features most predictive of VA were a greater peak count in V1, a greater # of peaks less <6.25% of the QRSd in V1 and a greater # of peaks in the final quarter of the QRS of V1 (Figure). ML models of peak amplitude, QRSd and fQRS were not predictive of VA. The addition of clinical variables to the QRS peak count, width and location models did not improve VA prediction. Arrhythmogenic ventricular substrate can be independently identified during intrinsic rhythm by QRS fragmentation in V1 characterized by either a greater total # of peaks, a greater # of very narrow peaks, or a greater # of peaks at the end of the QRS according to our ML model. These findings require external validation and may improve risk stratification in defibrillator-eligible CM patients.Tabled 1Table. Performance of Random Forest ML Model.ModelROC [95% CI]Prespecified SensitivitySpecificityPPVNPVF1-ScoreQRS duration0.57 [0.35-0.87]90%0.220.270.490.42Qualitative fQRS0.56 [0.33-0.73]90%0.120.240.350.38All clinical characteristics∗0.46 (0.22-0.70]90%0.220.260.930.41QRS peak characteristics- Total Count0.87 [0.65-1.00]90%0.640.501.000.64- Width0.88 [0.70-1.00]90%0.730.541.000.69- Amplitude0.66 [0.36-0.93]90%0.250.290.710.44- Temporal Location in QRS0.84 [0.66-1.00]90%0.630.471.000.62QRS peak + clinical characteristics*- Total Count0.79 [0.55-0.97]90%0.560.421.000.57- Width0.82 [0.58-0.97]90%0.630.451.000.61- Amplitude0.68 [0.42-0.93]90%0.360.320.960.48- Temporal Location in QRS0.76 [0.52-0.95]90%0.530.391.000.55Age, Sex, Cardiomyopathy, QRS duration, Heart Rate, LVEF, NYHA Class, Comorbidities and Cardiac Medications Open table in a new tab