A Signal Processing Approach for Detection of Hemodynamic Instability before Decompensation

Abstract

Advanced hemodynamic monitoring is a critical component of treatment in clinical situations where aggressive yet guided hemodynamic interventions are required in order to stabilize the patient and optimize outcomes. While there are many tools at a physician's disposal to monitor patients in a hospital setting, the reality is that none of these tools allow hi-fidelity assessment or continuous monitoring towards early detection of hemodynamic instability. We present an advanced automated analytical system which would act as a continuous monitoring and early warning mechanism that can indicate pending decompensation before traditional metrics can identify any clinical abnormality. This system computes novel features or bio-markers from both heart rate variability (HRV) as well as the morphology of the electrocardiogram (ECG). To compare their effectiveness, these features are compared with the standard HRV based bio-markers which are commonly used for hemodynamic assessment. This study utilized a unique database containing ECG waveforms from healthy volunteer subjects who underwent simulated hypovolemia under controlled experimental settings. A support vector machine was utilized to develop a model which predicts the stability or instability of the subjects. Results showed that the proposed novel set of features outperforms the traditional HRV features in predicting hemodynamic instability.

Conflict of interest statement

Competing Interests: The authors have read the journal’s policy and the following authors of this manuscript have the following competing interests: AB KN KW HD: Invention disclosure, (pending) patents pertinent to the product in development (on the algorithm described in this manuscript filed through the University of Michigan’s Office of Technology Transfer. The patent filing is currently under review: U.S. Application No. 14/751,260; Title: Early Detection of Hemodynamic Decompensation Using Taut-String Transformation"). The following authors have declared that no competing interests exist: MS SA VC. This does not alter the authors’ adherence to all PLOS ONE policies on sharing data and materials.

Figures

Fig 1. Heart Rate and Blood Pressure.

The central mark within each box is the median, the edges of the box are the 25th and 75th percentiles, the whiskers are the standard deviations of the data at each stage and the outliers are plotted in red ‘+’. (a): Box plot depicting the interpersonal variability of HR in each stage across all subject; (b): Box plot of overall mean and standard deviation of arterial BP across each stage of the LBNP experiment.

Fig 2. Methods Flowchart.

Outline of the various discrete steps involved in the designed methodology.

Fig 3. Taut String.

X axis is the time in seconds (example shown is 60 seconds), Y axis is epsilon width from the origin for piecewise linear reconstruction. (a)The black and red waveforms are the taut-string margins, z − ε and z + ε, and the dotted blue lines is the taut-string estimation; (b) Shows the reconstructed signal based on the taut-string estimation.

Fig 4. Top row.

the HRV signals and their taut string estimates for a window at baseline, middle stage and final stage. Bottom row: the differences between the HRV waveforms and their taut string estimates. The figure also shows the value of the Taut String 32 feature at each stage.

Fig 5. Predicted versus actual severity levels.

The predicted severity levels are computed using multiple linear regression of the 10 selected features versus the typical HRV features. The bars indicate 95% confidence intervals.

Fig 6. AUCs and Accuracies for Each Feature Set.

(a) Areas under the receiver operating curve (AUCs). For the presented features, pair-wise Mann-Whitney U tests were used to compare the AUC’s for different window sizes to the one with the highest AUC (210-beats). The differences that are statistically significant are denoted by asterisks. The same was done for the typical HRV features and window sizes whose AUC was significantly different from the best window size (180-beats) are denoted by plus signs; (b) Accuracies averaged across all ten validation folds for each window size. Error bars represent a 95% confidence interval.

Fig 7. ROC.

Receiver operating curves for each feature set’s highest-performing window size.

Fig 8. Featureset Sizes.

Area under the curve calculated choosing the top ten, eight, six, four, and two features from the presented features’ best-performing window size (120-beats). The AUC for the Typical HRV features’ best performing window size (150-beats) is included (in blue) for comparison. The presented features continue to outperform the entire typical HRV feature set even when reduced to just four of the original ten features. The error bars indicate 95% confidence intervals.