have_you_met_rahul
About me Research Blog FAQ
have_you_met_rahul

Do HRV right

  1. ECG Time Series Variability Analysis Engineering and Medicine by Cornforth, David Jelinek, Herbert Franz Khandoker, Ahsan Habib, 2018
  2. Considerations in the assessment of heart rate variability in biobehavioral research
  3. Simple HRV analysis pipeline and documentation from Neurokit2

Water intake increases HRV in a dose dependent manner Water consumption as a source of error in the measurement of heart rate variability Caffeine intake, some medications, alcohol consumption, time of day, exercise, food inatake, water intake and bladder filling influence HRV

  1. Statistical considerations for reporting and planning heart rate variability case-control studies
  2. Guidelines for Reporting Articles on Psychiatry and Heart rate variability (GRAPH): recommendations to advance research communication
Note
So, I collected ECG data using lead I and Neurokit2 failed to pickup the R peaks correctly and ended up picking up S! Deep diving to know why

Let us look at what happens to the ECG data once it enters the pipeline.

Stop 1: ecg_clean

  • Default method to clean ECG is neurokit
  • Check for any nan in the data and returns missing datapoints
  • If there are gaps, they are filled with last available value (forward filling method)
  • There is a high pass filter set at 0.5 Hz butterworth filter (order = 5), followed by a powerline filtering (By default, powerline = 50)

Stop 2: ecg_peaks

  • Default method to identify peaks is neurokit
  • QRS complexes are detected based on the steepness of the absolute gradient of the ECG signal. Subsequently, R-peaks are detected as local maxima in the QRS complexes. Refer this paper for details

In short

  1. Compute the ECG gradient to detect rapid changes in slope (rate of change)
    Take absolute values of gradient so that we don’t miss out negative going waves
  2. Smooth the absolute gradient to reduce noise using one tenth of a second long boxcar kernel
    One more level of smoothing using 75 percent of a second long kernel. This would be ultra smooth
    3.Set a dynamic threshold to identify QRS complexes using the double smooth signal. The threshold is set at 1.5 times the smooth curve optimised for finding peaks
  3. Detect QRS start and end points based on threshold crossing
  4. Filter out short QRS complexes that are likely noise The values are set at one third of a second
  5. Find the most prominent peak within each QRS complex. This happens via the locmax[np.argmax(props["prominences"])]. locmax sees only peaks and not troughs

This is really critical for R peak detection. If our ECG waveform is flipped (R peaks pointing down), the algorithm would pick the S peak (falling end of R) instead of the peak! See the final comparisons

  1. Apply a refractory period to prevent false detections. Once a peak is detected for identified QRS complex, it compares its proximity to previously identified peak of previous QRS complex. If it is within one third of a second, the current peak is dropped. More like a refractory period idea implemeneted
  2. See the identified peaks

See what happens when we flip the ECG waveform

R peaks pointing up

R peaks pointing down

Observations

  • The delay in R peak (time from R peak to S peak)
  • The whole interval gets shifted
  • The numbers above red dots in panel 1 are the locations of R peak detections
  • The algorithm becomes an S peak detector
  • Should we be worries? YES offcourse
  • The RR interval remains more or less similar in both the scenarios

Positive: 664, 680, 723, 732, 711, 707, 737, 738, 734, 713, 726, 752, 718

Negative: 664, 680, 722, 732, 712, 706, 737, 738, 734, 713, 726, 752, 719