Fig. 4: Temporal localization of ECG features associated with WMAs identified by ECG-WMA-Net. | npj Digital Medicine

Fig. 4: Temporal localization of ECG features associated with WMAs identified by ECG-WMA-Net.

From: Identification of cardiac wall motion abnormalities in diverse populations by deep learning of the electrocardiogram

Fig. 4

a Times of cardiac activation on the ECG that identify abnormal wall motion. Columns A, B, and C are the output of SHAP analysis of trained machine learning models to each of the patient cases from Fig. 2, in order. The cropped and aligned ECG signal from each patient is shown in red trace. The column labeled aggregate displays the summated ECG SHAP values output map for all patients. Both the example cases and the aggregate shadings support that ECG features throughout the QRS and T-waves were used to identify WMA. b A sliding window of 120 ms was used to build successive models for each 40 ms from 60 to 540 ms, while ablating the remainder of waveform data. The black line shows the ROC AUC values with confidence intervals for each model trained on windows centered at each time point. The red tracing indicates a standardized average ECG with normalized voltage.

Back to article page