Abstract
BACKGROUND AND OBJECTIVE: Despite the need for continuous, accurate, and minimally invasive cardiac output measurement in pediatric patients, no recommended methods currently exist. This study aimed to develop and validate a deep learning (DL)-based algorithm that utilizes a minimally invasive arterial pressure waveform to measure pediatric stroke volume index (SVI) accurately. METHODS: A total of 70 pediatric operations (on 67 patients) were included, with a median age of 6 years. We derived stroke volume (SV) from the Doppler-based stroke distance and aortic cross-sectional area, and then adjusted it to the SVI using body surface area. The arterial pressure waveform and patient demographics were used to predict SVI. The model was validated using error, Bland-Altman, four-quadrant plot, cycle-to-cycle variability, and sensitivity analyses. A saliency map was used to visualize the model’s comprehension of the waveform. RESULTS: The DL model demonstrated a mean absolute error of 4.1 ± 2.8 mL/m(2). The limit of agreement (LOA) ranged from − 11.23 ± 0.01 mL/m(2) to 8.7 ± 0.1 mL/m(2), which was ± 28.6% of bias. In the sensitivity and saliency map analyses, the model effectively extracted features from the arterial pressure waveform. CONCLUSIONS: The deep learning model showed promising accuracy (LOA < 30%) in estimating SVI in children. However, its ability to track rapid hemodynamic changes is limited in some cases. Further improvement will be useful for optimizing systemic oxygen delivery and enhancing patient outcomes in pediatric care. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12911-026-03428-x.