Fig. 5: Automatic image processing by deep learning.

a, Schematic workflow. Pre-processed images are used to train the FCN-32 model. The trained model accepts the unprocessed images and predicts the left ventricular (LV) volume, based on which stroke volume, cardiac output and ejection fraction are derived. b, Left ventricular volume waveform generated by the FCN-32 model from both the wearable imager (W.I.) and the commercial imager (C.I.) (left). Critical features are labelled in one detailed cardiac cycle (right). c, Bland–Altman analysis of the average of (x axis) and the difference between (y axis) the model-generated and manually labelled left ventricular volumes for the wearable (black) and commercial (red) imagers. Dashed lines indicate the 95% confidence interval and about 95% of the data points are within the interval for both imagers. Solid lines indicate mean differences. d, Comparing the stroke volume, cardiac output and ejection fraction extracted from results by the wearable and commercial imagers. Data are mean and s.d. from twelve cardiac cycles (n = 12). e, The model-generated left ventricular volume waveform in the recovery stage. f, Three representative sections of the recording from the initial, middle and end stages of e. End-systolic volume (ESV), end-diastolic volume (EDV), stroke volume and ejection fraction (g) and cardiac output and heart rate waveforms (h) derived from the left ventricular volume waveform. The end-systolic volume and end-diastolic volume gradually recover to the normal range in the end section. The stroke volume increases from about 60 ml to about 70 ml. The ejection fraction decreases from about 80% to about 60%. The cardiac output decreases from about 11 l min⁻¹ to about 9 l min⁻¹, indicating that the decreasing heart rate from about 175 bpm to about 130 bpm overshadowed the increasing stroke volume. AS, atrial systole; IC, isovolumetric contraction; IR, isovolumetric relaxation; RI, rapid inflow.