Estimation of Left Ventricular End-Systolic Elastance From Brachial Pressure Waveform via Deep Learning

Determination of left ventricular (LV) end-systolic elastance (Ees) is of utmost importance for assessing the cardiac systolic function and hemodynamical state in humans. Yet, the clinical use of Ees is not established due to the invasive nature and high costs of the existing measuring techniques. The objective of this study is to introduce a method to assess cardiac contractility, using as a sole measurement an arterial blood pressure (BP) waveform. Particularly, we aim to provide evidence on the potential in using the morphology of the brachial BP waveform and its time derivative for predicting LV Ees via convolution neural networks (CNNs). The requirement of a broad training dataset is addressed by the use of an in silico dataset (n = 3,748) which is generated by a validated one-dimensional mathematical model of the cardiovasculature. We evaluated two CNN configurations: 1) a one-channel CNN (CNN1) with only the raw brachial BP signal as an input, and 2) a two-channel CNN (CNN2) using as inputs both the brachial BP wave and its time derivative. Accurate predictions were yielded using both CNN configurations. For CNN1, Pearson’s correlation coefficient (r) and RMSE were equal to 0.86 and 0.27 mmHg/ml, respectively. The performance was found to be greatly improved for CNN2 (r = 0.97 and RMSE = 0.13 mmHg/ml). Moreover, all absolute errors from CNN2 were found to be less than 0.5 mmHg/ml. Importantly, the brachial BP wave appeared to be a promising source of information for estimating Ees. Predictions were found to be in good agreement with the reference Ees values over an extensive range of LV contractility values and loading conditions. Therefore, the proposed methodology could be easily transferred to the bedside and potentially facilitate the clinical use of Ees for monitoring the contractile state of the heart in the real-life setting.