Fig. 2: Quantitative relationships between eAPs and iAP waveform features.
a A representative simultaneous recording of eAP and iAP from arrhythmic cells (top) demonstrates a strong correlation (r = 0.90) between eAP amplitude (maximum voltage during the spiking phase [mV]) and iAP spike velocity (percentage change in iAP voltage over time [% change in iAP/s]). This association is further evident as oscillations in the extracellular recordings reflect the action potential’s repolarization phase. b Key descriptors to characterize eAP and iAP waveforms and examples of distorted eAPs. The details are described in the Methods section. c A correlation analysis between eAP features and iAP APD values on undistorted eAP/iAP pairs (n = 1049). d The distributions of normalized eAP features alongside iAP features (n training = 1512, n Test1 = 272, n Test2 = 171 and, n Test3 = 91). e Examples of XGBoost predicted and actual APD lines for the three test sets. f The distribution of prediction errors for APD values across the training set and test sets (n training-val = 1209, n Test1 = 272, n Test2 = 171 and n Test3 = 91). g Distribution of mean APD errors for Test1 (0.020 ± .007 s), Test2 (0.040 ± .006 s), Test3 (0.047 ± .038 s), and the training set (0.002 ± .001 s) for n training-val = 1209, n Test1 = 272, n Test2 = 171 and, n Test3 = 91. h SHAP values identify ΔTd, ΔV1/ΔV2, ΔTs, ΔV1, and IR as the most significant eAP features for predicting iAP features. SHAP values illustrate how varying these features affects the predicted values relative to the model’s average output, with feature importance defined as the average of absolute changes imposed on predictions by varying feature values. i The ranked significance of eAP signal features in predicting APD values according to partial dependency plots. Each dot represents a single predicted APD value, with its color indicating the feature’s value and its position on the X-axis showing its SHAP value, reflecting the expected deviation in APD prediction. For example, in the APD90 plot, increasing ΔTs (transitioning from blue to red) results in higher predicted APD90 values, whereas ΔTs and APD30 exhibit the opposite trend. Note that SHAP values may be influenced by local minima, potentially limiting their representation of the global relationship between features. The box plots show the median (center line), interquartile range (IQR; box bounds), whiskers (1.5 × IQR), and outliers (points beyond whiskers).
