Fig. 4: Prediction of BP with correction for hydrostatic pressure error.

A Box plots of h-stratified MAE for DBP (middle) and SBP (bottom) predictions generated using the uncorrected and corrected model compared to the measured reference across 20 subjects with n = 19, 20, and 19 measurements for heights of h = −25, 0, and 25 cm, respectively. Each measurement was the participant’s time-averaged MAE for the indicated h. The box represents the interquartile range, with the horizontal line at the median value. The vertical lines extend to the maximum and minimum data points within 1.5 × IQR. The diagrams (top) show an illustrative example of an arm pose corresponding to each group. Significance determined by mixed-effects model followed by Šídák post hoc test. (DBP: for h = -25 cm, ****P < 0.0001; for h = 0 cm, *P = 0.0356; for h = 25 cm, **P = 0.0096. SBP: for h = -25 cm, ****P < 0.0001; for h = 0 cm, *P = 0.0284; for h = 25 cm, **P = 0.0051). Individual trajectories in Supplementary Figs. 13 and 14. B, C Repeated measures Bland-Altman plots using the uncorrected and corrected model compared to the measured reference for DBP (D) and SBP (E) prediction across time-averaged participant and height pairs (n = 58; 2 to 3 measurements from each of the 20 participants). The X-axis shows the average of prediction and reference, and the y-axis shows the difference between prediction and reference. The solid line indicates the mean difference, and the dashed line indicates 95% LoA. Enlarged versions showing the values of h in Supplementary Figs. 15,16. D A representative participant’s time series of measured and predicted hand height (top); corresponding time series of DBP and SBP predicted using the uncorrected and corrected models compared to the measured reference (bottom). Time series of representative participant’s BP prediction error in Supplementary Fig. 17. E Participant-aggregated MAE over time for DBP (top) and SBP (bottom) predictions using the uncorrected and corrected model. The first 4 min of data were used for calibrating the models, while the following 8 min were used for prediction. Data were represented as mean ± standard error of the mean (n = 13 to 20 for each point, across the 20 subjects).