We performed a quality check for all impedance cardiography datasets, excluding measurement errors defined as abnormal changes in heart rate (change > 30% in less than 3 s). In line with manufacturer recommendations, we averaged the beat-to-beat data over four heartbeats.
Blood pressure measurements were assessed for undisturbed pressure curves. We excluded measurements when there was an increase of cuff pressure of more than 8 mmHg during cuff deflation.
To decide whether to use the ECG’s Q-wave (physiological beginning of PEP = start of ventricular depolarization) or the R-wave (easier detection), we compared the PEP derived from both starting points using a correlation analysis and analyzed the changes in the Q-R-time and its dependency to mental or physical load.
We analyzed changes of PEP during the TSST and ergometer load separately by modeling a mixed linear model (IBM SPSS Statistics 27). For Regression analyses, we performed linear regression and non-linear, R2-calculation via Scikit-learn (30 ). To find differences in PEP’s behavior under different circumstances, we calculated a mixed linear model and adjusted for the heart rate as a covariate factor. Subsequently, we trained a k-nearest-neighbor classifier with patient mean values for rest, TSST and bike ergometer load. In accordance with best practices of Machine Learning we performed a strict separation of training and test data. We tested the classifier using a subject-dependent train-test split. This means that all three data-points of any given patient had to be either in the train or the test group, therefore reducing the change for data-leakage. We retrieved our results by evaluating the classifier in an 80/20 k-fold (equates to fivefold) evaluation scheme. We averaged the results (positive predictive value/sensitivity) over all three outcomes (rest, mental load, physical load) and over all k-fold iterations to provide an aggregate estimation of the classifiers performance.
To illustrate patient specific differences at rest and under mental and physical load, we visualized data at rest measurement, first TSST question and last ergometer step (maximum load) in a boxplot. We analyzed the differences via a one-way ANOVA. The statistical analysis was performed in close collaboration with Charité’s Institute of Biometry and Clinical Epidemiology. We used the conservative Bonferroni correction for all multi-group comparisons.
To estimate the effect of neglecting or estimating the PEP, we analyzed a previously published and often-cited relation between PWV and BP (6 (link)). We applied the relation for a standard human (1.80 m) and translated the relation to a pulse-arrival-time (PAT, time from ECG Q-wave to arrival of the pulse wave in the periphery) vs. BP relation. Thereafter, we used our data to estimate the PEP’s proportion of the PAT for different BP levels (regression) and subtracted the PEP to receive the pulse-transit-time (PTT = PAT–PEP). We then analyzed the PEP estimation uncertainty (intra- and interindividual PEP variability at similar BP levels) and calculated confidence intervals for one SD (67% of values) and two SDs (95% of values) for PTT. Applying this relation, we were able to determine the BP measurement estimation uncertainty caused by either neglecting or estimating the PEP for PWV based BP measurement.
Blood pressure measurements were assessed for undisturbed pressure curves. We excluded measurements when there was an increase of cuff pressure of more than 8 mmHg during cuff deflation.
To decide whether to use the ECG’s Q-wave (physiological beginning of PEP = start of ventricular depolarization) or the R-wave (easier detection), we compared the PEP derived from both starting points using a correlation analysis and analyzed the changes in the Q-R-time and its dependency to mental or physical load.
We analyzed changes of PEP during the TSST and ergometer load separately by modeling a mixed linear model (IBM SPSS Statistics 27). For Regression analyses, we performed linear regression and non-linear, R2-calculation via Scikit-learn (30 ). To find differences in PEP’s behavior under different circumstances, we calculated a mixed linear model and adjusted for the heart rate as a covariate factor. Subsequently, we trained a k-nearest-neighbor classifier with patient mean values for rest, TSST and bike ergometer load. In accordance with best practices of Machine Learning we performed a strict separation of training and test data. We tested the classifier using a subject-dependent train-test split. This means that all three data-points of any given patient had to be either in the train or the test group, therefore reducing the change for data-leakage. We retrieved our results by evaluating the classifier in an 80/20 k-fold (equates to fivefold) evaluation scheme. We averaged the results (positive predictive value/sensitivity) over all three outcomes (rest, mental load, physical load) and over all k-fold iterations to provide an aggregate estimation of the classifiers performance.
To illustrate patient specific differences at rest and under mental and physical load, we visualized data at rest measurement, first TSST question and last ergometer step (maximum load) in a boxplot. We analyzed the differences via a one-way ANOVA. The statistical analysis was performed in close collaboration with Charité’s Institute of Biometry and Clinical Epidemiology. We used the conservative Bonferroni correction for all multi-group comparisons.
To estimate the effect of neglecting or estimating the PEP, we analyzed a previously published and often-cited relation between PWV and BP (6 (link)). We applied the relation for a standard human (1.80 m) and translated the relation to a pulse-arrival-time (PAT, time from ECG Q-wave to arrival of the pulse wave in the periphery) vs. BP relation. Thereafter, we used our data to estimate the PEP’s proportion of the PAT for different BP levels (regression) and subtracted the PEP to receive the pulse-transit-time (PTT = PAT–PEP). We then analyzed the PEP estimation uncertainty (intra- and interindividual PEP variability at similar BP levels) and calculated confidence intervals for one SD (67% of values) and two SDs (95% of values) for PTT. Applying this relation, we were able to determine the BP measurement estimation uncertainty caused by either neglecting or estimating the PEP for PWV based BP measurement.
Free full text: Click here
