RapidHRV: an open-source toolbox for extracting heart rate and heart rate variability

Figure 1. Overview of RapidHRV pipeline. Across an entire block, the pipeline initially processes data with high-pass filtering, upsampling, and smoothing. RapidHRV then applies a sliding window across the entire block. Within each window, the data are scaled. Heart rate (beats per minute) and heart rate variability (root mean squared of successive differences + standard deviation of intervals) are calculated for each window, and data is submitted to outlier rejection. RapidHRV produces both a cleaned and uncleaned time series of heart rate and heart rate variability.

Figure 2. Example of how the implementation of traditional amplitude-based analyses may be insufficient for peak detection, particularly for atypical ECG signals. Traditional amplitude-based peak detection (top) incorrectly identifies T waves as the relevant peak of interest. However, using RapidHRV’s implementation of k-means clustering (bottom), we are able to derive measures based on the correct peaks (i.e., R waves).

Figure 3. Outlier rejection methods for Rapid HRV. The only outlier rejection method applied to the uncleaned time series is screening for a sufficient number of peaks to derive metrics. The cleaned time series then goes through a battery of biological constraints (thresholding minimum/maximum beats per minute (BPM) and root mean square of successive differences (RMSSD)) and statistical constraints (median absolute deviation (MAD) of peak heights, prominences, and intervals; ensuring adequate duration from first to last peaks).

Figure 4. Interactive visualization. The user is presented with the timeseries of HR/HRV metric (specific measure can be selected from the top tab; i.e., RMSSD has been selected). Users can click on specific timepoints to view extraction windows (top right; foreground).

Figure 5. Parameter recovery of simulated PPG data as a function of heart rate variability. Parameter recovery of simulated PPG data as a function of heart rate variability (Intraclass Correlations (ICC) and Root Mean Square Error (RMSE) against ground truth). Y axes reflect the true RMSSD in the data, whilst the X axes reflect RapidHRV’s estimation. For readability, data is only plotted in a key range of performance (subplots: 10–40 dB noise left-to-right; 20–250 Hz top-to-bottom).

Figure 6. Average estimated effect size of simulated ‘anxiety’. Average estimated effect size (+95% Confidence Intervals) of simulated ‘anxiety’ on the root mean square of successive differences as a function of signal to noise ratio (SNR), sampling rate, and cleaning. Higher values indicate greater sensitivity. This suggested RapidHRV was able to estimate true effect sizes in data when signal to noise ratio was >=30 dB. Visualized with ggplo2 (Wickham, 2016).

Figure 7. Effects of age on RapidHRV-estimated RMSSD. RainCloud plot (Allen et al., 2019) demonstrating effects of age on RapidHRV-estimated RMSSD (cleaned) derived from movie-watching ECG (N = 40). As expected, older participants typically had lower heart rate variability than younger participants.

Figure 8. Agreement between RapidHRV and a previous Labchart analysis. Agreement between RapidHRV and a previous Labchart analysis (de Groot, Kirk & Gottfried, 2020) of heart rate and variability in a finger IR PPG dataset (N = 39).

Figure 9. Motion across wrist PPG activities. RainCloud plot (Allen et al., 2019) of motion across activities. Each datapoint represents an individual subject’s mean momentary jerk magnitude within a condition. Curves demonstrate distribution density across the sample (kernel density plot).

Figure 10. Agreement between PPG-based RapidHRV-estimates and visually inspected ECG-based estimates across different activities.