Measuring autonomic activity Heart rate variability
Centre for Doctoral Training in Healthcare Innovation
Dr. Gari D. Clifford, University Lecturer & Director, Centre for Doctoral Training in Healthcare Innovation, Institute of Biomedical Engineering, University of Oxford Autonomic regulation
HRV metrics . Time domain . Spectral . Nonstationary ▪ Short term – PRSA, HRT ▪ Long term – wavelet scaling
Dealing with noise in the time series . Resampling issues . Removing abnormal beats
ECG-derived respiration (EDR) . Physical . Autonomic
Rest & Digest Fight & Flight
http://www.becomehealthynow.com/images/organs/nervous/sympth_parasymth.gif ANS autoregulates heart through SA node
So measure HRV to gain insight into how autoregulation is working
Why? . To provides a metric of health . … by looking for departures form normality (given demographics) Sequence of RR intervals is called Tachogram 60/.RR interval = Instantaneous HR Plot (t,RR) (time vs differential of time!) Now you can employ signal processing on the data!
Data taken from PhysioNet; http://www.physionet.org Oscillations in RR tachogram from:
Simple example: RSA
Also changes due to blood pressure
Myogenic changes? Smooth muscle
Diurnal variations (temp, sleep, activity)
Sudden changes – ectopy, arrhythmia
ECG, HR and Respiration:
Resp rate is highest HR and respiration highly correlated Freq component ~ 0.1-0.5 Hz
HRV is a quantification of variation of the beat-to-beat intervals
Frequency domain analysis is traditionally split into 4 frequency bands (ULF, VLF, LF & HF) representing the 4 (approx. distinct) time scales over which cardiovascular variations are thought to occur
• HF (0.15-0.4Hz): • VLF (0.003-0.04Hz): Myogenic? Vagal/parasympathetic variations Variations over hours, e.g. temp. over seconds (e.g. respiration) • ULF (0.0001-0.003Hz):Circadian • LF (0.04–0.15Hz): Sympathetic- – e.g. activity nonstationarities over minutes (BP, Meyer waves) Sympathetic & parasympathetic braches of CNS act in opposition Think of it like 2 pedals in a car – both are accelerators AND breaks (innervate / inhibit) The sympathetic brake/accelerator is less ‘sticky’ Fight/flight response = rapid sympathetic innervation, and slower parasympathetic inhibition (note that parasympathetic action leads to higher frequency oscillations in RR tachogram – why?)
The ratio of the LF & HF power reflects the LF HF ‘autonomic balance’ between these continuously interacting inhibitory and innervating actions of the CNS
Small values indicate you are relaxing Large values indicate a highly active system – e.g. when you are running Elevated values when you are relaxing indicate health problem Record ECG
Detect Peaks
Calculate RR interval (time between each R peak)
Remove RR intervals associated with noise & non-sinus beats (& following N beats?)
Interpolate through missing data (insert phantom beats) if using cubic spline? [or cut data] (why do we remove non-sinus beats?)
Resample / interpolate time series – WHY?
Data is unevenly sampled!
Nyquist not strictly defined
Data is unevenly sampled!
Nyquist not strictly defined
Nyquist frequency = 1/mean(RR) Hz (generally 0.5Hz for 60 BPM)
Some frequencies above this, so you can ‘beat’ Nyquist through uneven sampling
But the accuracy at higher frequencies depends on number of samples with corresponding intervals
Data courtesy of PhysioNet; http://www.physionet.org Non-sinus beats are not reflective of SA node activity They lead to nonstationarities in time series … so we remove a few following beats Insert phantom beats to create stability for nonlinear interp Resample to make an evenly sampled time series
Generally you remove any RR intervals which change by more than 20% on the previous RR interval
Example with linear interpolation
Note phantom beat is interpolated with linear interp – so phantom not needed here
Depends on task.
Is (average) Nyquist the right resampling frequency?
Smallest RR interval possible? → ? Hz
For respiration, what’s the fastest rate? → ? Hz
Autonomic information up to ? Hz
Depends on task.
There are frequencies above average Nyquist!
Smallest RR interval possible? 200ms → 10 Hz
But HRV only for sinus beats: <160 BPM → ~6 Hz
Fastest respiration rate? 60 RPM → 2 Hz
Autonomic information up to 1 Hz → 2 Hz
(In reality you need to sample a bit faster than Nyquist) Many measures, some parametric, some non- parametric.
Work over many scales
Some deal with non-stationarities, some do not Sine wave, with standard deviation calculated over different lengths
Eventually it tends to a limit, but local measures can be unrepresentative
Time domain (assumes stationarity)
Standard frequency domain (assumes stationarity)
‘Scaling’ over 24 hours
Nonlinear measures (sample entropy)
Multi-scale entropy
Recap: Moments of a distribution Discrete approx Continuous
Mean
Var
General:
x Gaussians are mesokurtic with κ =3
SubGaussian SuperGaussian
x Approximations of distributions:
AVNN: Average of all NN intervals
SDNN: Standard deviation of all NN intervals
SDANN: Standard deviation of the averages of NN intervals in all 5-minute segments of a 24-hour recording
SDNNIDX: Mean of the standard deviations of NN intervals in all 5-minute segments of a 24-hour recording
rMSSD: Square root of the mean of the squares of differences between adjacent NN intervals
pNN50: Percentage of differences between adjacent NN intervals that are > 50 ms. This is one member of the larger pNNx family LF, HF, LF/HF, VLF, ULF
Why these bands?
Chemical experiments provide evidence …
. HF (0.15-0.4Hz): Vagal/parasympathetic variations over seconds (e.g. respiration)
. LF (0.04–0.15Hz): Sympathetic- over minutes (BP, Meyer waves)
. VLF (0.003-0.04Hz): Myogenic? Variations over hours, e.g. temp.
. ULF (0.0001-0.003Hz):Circadian – e.g. activity nonstationarities
Tachogram has many states with HRi similar means or variances
Length of state varies minutes (weakly stationary)
Movements between states have brief accelerations in RR interval new mean and/or variance.
Time (40 mins) 24 hour spectrum of RR intervals exhibits 1/f β scaling
β indicates type of long term correlation . β=2 : Brownian motion . β=1 : Pink (natural) noise . β=0 : White noise (no long-term correlation) . Measure of ‘fractal’ properties? . Scaling should be not too white, or too brown. Pink is normal for humans (Probably not)
How do we measure β since HRV is not stationary? Hint: not with Fourier.
Entropy kln(W) is a measure of disorder … the more random the time series, the more disorder
HRV should have some randomness, but not too much. (c.f. scaling)
More info at: http://physionet.org/physiotools/ApEn/
http://sdic.sookmyung.ac.kr/pharmacotherapy/INSOM/sleep_cycle.jpg
Only short segments of data required Unit-free - no scaling issues Thought to reflect the sympathovagal balance
HRV changes Wakefulness Deep Sleep significantly in different sleep cycles and for different conditions: REM (Dream) Sleep Light Sleep Activity : LF/HF ratio :
Spontaneous Breathing 1.39 0.28
Controlled Breathing (15 rpm) 0.69 0.37
Controlled Breathing (18 rpm) 1.09 0.36
Silent Reading 1.52 0.26
Reading Aloud 1.59 0.21
Free Talking 3.58 0.45
Performing mentally stressful tasks silently (e.g. arithmetic) 3.05 0.39
Performing mentally stressful tasks aloud 2.89 0.31
Changes can be larger than inter-patient differences with different pathologies (Bernardi et al.)
HRT – Heart rate turbulence – a cardiovascular response to ectopy
PRSA – Phase rectified signal averaging – the normal response of HR accelerations and decelerations Data courtesy of PhysioNet; http://www.physionet.org
SA node response to ectopic beat; short HR acceleration then deceleration. Maintain BP; rapid parasympathetic withdrawal? Then parasympathetic innervation baseline
http://www.h-r-t.org/hrt/en/hrtdemo_js.html Credit: R. Schneider: http://www.librasch.org/
Ectopic beats disturb RR tachogram stationarity
Disturbance lasts 10 - 20 beats
HRT quantifies this disturbance using 2 metrics:
. TO: Turbulence Onset
. TS: Turbulence Slope
Credit: Bauer A, Barthel P, Schneider R, Schmidt G. Dynamics of Heart Rate Turbulence. Circulation 2001b; Vol. 104; No. 17; Supplement; II-339, 1622. (+ index intervals after ectopic, - index before)
Percentage difference between mean of each pair of NN intervals on either side of ectopic pair
Must average the TO over >> 10 ectopics Find steepest slope for each possible sequence of 5 consecutive normal intervals from RR+2 RR+16 Usually average 10- 20 time series first then calculate one TS on the average time series!
Outlier Rejection Important: (See Notes) Run: http://www.librasch.org/hrt/en/hrtdemo_java.html
Figure Credit: Mäkikallio et al., Eur. Heart J., April 2005; 26: TO < 0 and TS > 2.5 are normal
(a healthy response to PVCs is a strong sinus acceleration followed by a rapid deceleration) http://www.librasch.org/prsa/en/
. An independent predictor of late mortality after acute MI [Schmidt 1999, Ghuran 2002, Wichterle 2004, Watanabe 2005, Baur 2006]
. Abnormal HRT Predicts Initiation of Ventricular Arrhythmias [Iwasa 2005]
. HRT indices appear to correlate better with EF than SDNN in Chagas disease [Tundo2005]
. HRT Predicts Cardiac Death in Patients Undergoing CABG [Cygankiewicz 2003]
. Prognostic Marker in Patients with Chronic Heart Failure [Kayama 2002]
. Risk Predictors in Patients With Diabetes Mellitus [Barthel 2000 Barthel 2002]
. LF/HF ratio indicates stress [Healey 2002 + others]
. LF/HF ratio separates normal and sleep apneoic patients (with sleep stage) [Clifford 2003]
. LF/HF ratio used to screen patients who are responsive to sleep treatments [Campana 2011]
HRV is scale dependent – short and long term metrics exist
Some metrics appropriate only to short scales
Can deal with nonstationarities by using: . Perturbation analysis . Signal averaging . Short term ‘boxing’ . Nonstationary measures e.g. wavelets.
Can also measure how HRV changes over scale
Must be careful to remove noise from data first
Choose resample rate based on task! (Always true of all pre-processing – there’s no ‘magic’ pre-processing unit)
Try to use a smooth resampler if you must
… or avoid resampling (advanced)
HRV very useful in risk stratification, but often need to use some other measure – univariate analysis is limited
SNA: Sympathetic Nerve Activity (recorded from peroneal nerve)
What if there is noise in the data (or ectopy)? Abnormal or Non-sinus beats (ectopics) generally appear earlier (or sometimes later) than when a sinus beat is expected
Ectopic beats usually replace the sinus beat and are followed by a compensatory pause. Their frequency is on average about 1 per hour for the NSRDB
Artefacts occur are additional and occur independently of the phase of the sinus rhythm with a recording method-dependent distribution.
Linear/cubic interpolation of RR intervals then perform FFT Over-estimation of LF and under-estimation of HF Use Lomb-Scargle periodogram to avoid interpolation Spectral estimation of unevenly sampled data without resampling
Variable integration step size
Equivalent to least squares fitting of sines to data!
Scargle, J. D. (1982). "Studies in astronomical time series analysis. II - Statistical aspects of spectral analysis of unevenly spaced data". Astrophysical Journal 263: 835. doi:10.1086/160554