Algorithm for Reliable Detection of Pulse Onsets in Cerebral Blood Flow Velocity Signals

Figures

• 

  Figure 1. Shown are samples taken from the data set. (A) Shows a good quality, healthy waveform for reference. (B) Demonstrates a common measurement error that results in signal with a digitized appearance. (C) Shows a signal with a significant high frequency noise component. (D) shows an example of a blunted waveform. (E) Shows a waveform with an elevated P2. (F) Shows a potentially stenotic waveform. (G) Demonstrates a signal with extremely poor quality. And finally, (H) illustrates a signal with a sharp noise spike.

• 

  Figure 2. The flowchart showing a high level overview of the steps involved in identifying beat onsets.

• 

  Figure 3. Shown are the major steps in the algorithm, from an initial seed onset to detection of the next onset. The x (time) and y (CBFV/MDF amplitude) axes have been normalized to dimensionless quantities for ease of viewing. The algorithm proceeds from onset to onset in this manner until the end of the scan is reached. (a) Shows the initial onset in the CBFV signal, which consists of the onset point itself and its associated peak. (b) Displays the MDF signal associated with this same segment of data and shows how the search window is established after enforcing the refractory period and locating the threshold crossing point. (c) Shows the CBFV signal with the search window overlaid. (d) Shows a zoom-in of the search window and the location of all valleys, with the new onset labeled. Finally, (e) shows the original CBFV signal with the new onset identified.

• 

  Figure 4. Shown above is a long beat resulting from a missed onset due to a suppressed upslope. The x (time) and y (CBFV/MDF amplitude) axes have been normalized to dimensionless quantities. The CBFV signal is displayed in (A), while the MDF signal is shown in (B). Also shown in (B) are the threshold values used for onset detection, which are incrementally relaxed until either an onset is found or until some minimum threshold value is reached. The search window for the new onset is indicated by the vertical dashed lines once the refractory period τ has been enforced.

• 

  Figure 5. Shown above is an example of the detection of a short beat. The short beat is shown in the main figure between the dashed vertical lines between its two neighboring beats. Detected onsets are marked by blue dots. The lengths lbefore, lshort, and lafter are also labeled. The mean beat calculated for this scan is also displayed in the subfigure shown in the upper right corner (not to scale), and its length lmedian is displayed correctly to scale alongside the other distances in the main figure. The x (time) and y (CBFV/MDF amplitude) axes have been normalized to dimensionless quantities.

• 

  Figure 6. Shown are the different types of classification errors. These are shown for illustrative purposes and are not actual output of the algorithm. The dotted vertical line indicates the correct position of the beat onset. (A) Shows the typical situation in which a false negative arises: the algorithm fails to mark an onset for an obvious beat in between two neighboring detected beats. (B) Shows an example of a false positive, in which an extra onset unassociated with any true onset has been detected. (C) Illustrates a misaligned beat in which a true onset has been found, but its placement is offset from the true location.

• 

  Figure 7. Shown in (A,B) are the false positives detected by the algorithm in the data set. The correctly identified manually annotated beat onsets are marked by filled-in blue dots. The false positives have been aligned on the dotted line and are marked by empty circles.

• 

  Figure 8. Shown in (A,B) are the false negatives detected by the algorithm in the data set. The correctly identified manually annotated beat onsets are marked by filled-in blue dots. The locations in which an annotated onset was indicated but no onset detected are marked by the dotted line.

• 

  Figure 9. Shown here is the relationship between the number of true positives and their temporal offset error.