Individual Recognition by Heart Rate Variability of Two Different Autonomic Profiles Related to Posture
Background Power spectrum analysis of heart rate variability (HRV) can estimate the state of sympathovagal balance modulating sinus node activity. In view of the large distribution of spectral variables, a recognition of well-defined physiological conditions has never been attempted on an individual basis.
Methods and Results We considered 10 spectral variables extracted from short segments (200 to 500 cardiac cycles) of 350 ECG tracings recorded in normal subjects in both supine and upright positions (700 patterns). The tracings were first ordered consecutively and subsequently assigned alternatively to a training or to a test set (each consisting of 175 cases, providing 350 patterns considered to be independent). A forecasting linear method estimated a normalized activation index (ranging from −1 for supine to +1 for upright) that concentrated the information derived from spectral variables and that identified, in the test set, individual by individual, ≈84% of corresponding body postures.
Conclusions The combined use of spectral methodology and forecasting analysis has revealed an information content embedded, per se, in a short series of RR intervals capable of recognizing, individual by individual, two different autonomic profiles related to posture.
Heart rate, in physiological conditions, is the result of sinus node pacemaker rhythmicity and, mainly, of sympathetic and vagal activities. The hypothesis that the power spectrum analysis of HRV might furnish valuable estimates of sympathetic and vagal modulations of RR interval by means of the normalized powers of the LF and HF components1,2 has recently been validated by experiments with graded passive tilt3,4 and with direct recordings of muscle sympathetic nerve activity during stepwise pressure changes.5 Accordingly, the reciprocal relationship between LF and HF seems to assess the sympathovagal balance.
However, power spectrum variables are characterized by large intersubject and intrasubject variations6 probably reflecting the dynamics of physiological control mechanisms over time, even during steady-state conditions.2,6,7 Therefore, previous studies have addressed only the differences between given data sets. The aim of this study was instead to recognize, individual by individual and independently from each other, the spectral profiles corresponding to two easily reproducible physiological conditions such as supine or upright posture.
The study included 350 healthy subjects (230 from previous studies and 120 new subjects; 161 women, 189 men; mean age, 35±17 years) from whom ECG and respiratory recordings were obtained in controlled laboratory conditions, as previously described.1,4 Each subject was studied in both the supine and upright positions (obtained either with active standing (n=171) or passive tilt to 90° (n=179).
Autoregressive Spectrum Analysis
Autoregressive spectrum analysis of the variability of a time series of ≈200 to 500 RR intervals can extract, as already described,1,2,6,7 the following 10 variables: RR interval, total power (or variance), LF and HF components in absolute units (ms2), the very-low-frequency component (<0.03 Hz), and the center frequencies of LF and HF. In addition, the LF/HF ratio and LF and HF in nu can be calculated.1,2,6,7 In all cases, the respiratory frequency was >0.15 Hz.8
To test the possibility of predicting the posture (either supine or upright) of the subjects, we used 50% of the data (training set) to calculate the mathematical forecasting model that was subsequently validated in the remaining 50% of the data (test set).
Individual data were ordered consecutively in their historical sequence, and subsequently, odd and even rank positions were assigned to the training or test set, respectively. In this way, this study does not suffer from seasonal effect and the sets are well matched for sex, age, and active or passive upright posture.
Training and test sets each held 350 patterns characterized by 10 power spectrum variables belonging to 175 subjects studied in both the supine and upright positions. The features related to both postures were considered to be independent.
We used the method of Madansky9 to describe the changes induced in each variable by two different postures (as targets, coded +1 for the upright and −1 for the supine posture). This approach was preferred to standard linear methods because it does not require a gaussian distribution of variables (and, in fact, variables such as variance, LF, and HF have a skewed distribution). The linear coefficients of each variable (ie, the weights) were used to calculate, for each pattern, an AI (that, for graphical purposes, was normalized from −1 for supine to +1 for upright posture) as follows: where AIi is the AI of the ith pattern, w0 is the threshold, k is the number of variables of interest, wj is the weight of the jth variable, and vij is the value of the jth variable of the ith pattern. The weights (w) were first calculated from the training set and subsequently used to estimate the AI for the remaining 175 supine and 175 upright patterns of the test set. In this test set, the ith pattern was correctly attributed when the signs of its target and of its AIi were identical.
A stable forecasting should correctly attribute more than half of the unknown cases, and the recognition rates of both supine and upright patterns should be almost equal.
An example of the 10 power spectrum variables studied in both supine and upright posture is illustrated in Fig 1⇓.
During the training set, the algorithm had to match the target, ie, the posture, which was classified by the experimenter, with the information that could be extracted from the interaction of the variables of interest. A pattern was correctly discriminated when target and AI signs were identical. During the test set, as well, a negative value of the AI was intended to recognize the supine and a positive value the upright position.
Such a blind forecasting on the test set was capable of correctly assigning 83.4% (146 of 175) of features to the supine group and 86.3% (151 of 175) to the upright group when 10 variables were evaluated simultaneously.
To identify the best subset of variables, all the possible combinations using from 2 to 10 variables were calculated for both sets, and the best results are reported in the Table⇓.
Three variables (RR, LFnu, and HFnu) were found to hold almost all the information content and could recognize 294 of 350 (84.0%) patterns overall, with a similarly good performance in both supine and upright groups. This is illustrated by Fig 2⇓, evidencing the AI of each individual feature belonging to either the training or test set. In this case, the weights of the variables were RR=−9.4×10−6, LFnu=7.4×10−5, and HFnu=−7.0×10−5, with a threshold=4.5×10−3. When one of these three variables was not considered, the forecasting provided inconsistent results. On the other hand, other combinations required at least four variables to furnish good discriminations and recognitions—in some cases, however, with more unbalanced results (see Table⇑). In particular, adding the variance to RR, LFnu, and HFnu did not improve the recognition. In this case the weights of the variables were RR=−1.1×10−5, LFnu=7.2×10−5, HFnu=−7.5×10−5, and variance=2.0×10−7, with a threshold=5.1×10−3. The use of the absolute values of LF and HF caused an unbalanced forecasting, impairing the recognition of supine group patterns (Table⇑). Finally, when further information was added to the model based on 10 variables, for example the age of the subjects, the recognition power of AI in the test set was slightly increased to 84.3% in the supine group and to 87.6% in the upright group.
In this study we report, for the first time, that a physiological noninvasive recording, such as the ECG, that reflects neural regulatory mechanisms contains intrinsic information that can be used to recognize, individual by individual, two different postures.
Supine and upright postures, as an example of easily reproducible physiological conditions, engender distinct levels of sympathetic and vagal activity and hence of sympathovagal balance. Small studies clearly indicated that indices obtained from spectral analysis of RR interval variability can be used to infer the changes in autonomic balance produced by progressive tilt3,4 or graded changes in arterial pressure.5
The large distribution of spectral variables, however, has allowed the description of differences only between conditions or various populations. The present study demonstrates that by concentrating all the information derived from spectral analysis in a single index, obtained with a forecasting method, an individual recognition can be attempted. This accomplishment was allowed by the progressive development of a large data set, as required by the mathematical procedures involved.
In addition to the RR interval, the most powerful variables for both discrimination and recognition rates seem to be LFnu and HFnu. Because these variables appear to be the most sensitive markers of sympathetic and vagal modulations, respectively,1–7,10 this study confirms their value on an individual basis as well. Admittedly, our data refer to two well-defined physiological conditions, in one of which, the upright posture, there is a clear release from baroreceptor restraint of sympathetic activity with its LF rhythmicity, together with vagal withdrawal.11 It remains to be investigated whether this or similar approaches might be appropriate to explore the continuum of cardiovascular pathophysiology.
Selected Abbreviations and Acronyms
|HRV||=||heart rate variability|
- Received August 28, 1997.
- Revision received September 23, 1997.
- Accepted October 13, 1997.
- Copyright © 1997 by American Heart Association
Pagani M, Lombardi F, Guzzetti S, Rimoldi O, Furlan R, Pizzinelli P, Sandrone G, Malfatto G, Dell’Orto S, Piccaluga E, Turiel M, Baselli G, Cerutti S, Malliani A. Power spectral analysis of heart rate and arterial pressure variabilities as a marker of sympathovagal interaction in man and conscious dog. Circ Res. 1986;59:178–193.
Malliani A, Pagani M, Lombardi F, Cerutti S. Cardiovascular neural regulation explored in the frequency domain. Circulation. 1991;84:482–492.
Bootsma M, Swenne CA, Van Bolhuis HH, Chang PC, Cats VM, Bruschke AVG. Heart rate and heart rate variability as indexes of sympathovagal balance. Am J Physiol. 1994;266:H1565–H1571.
Montano N, Gnecchi Ruscone T, Porta A, Lombardi F, Pagani M, Malliani A. Power spectrum analysis of heart rate variability to assess the changes in sympathovagal balance during graded orthostatic tilt. Circulation. 1994;90:1826–1831.
Pagani M, Montano N, Porta A, Malliani A, Abboud FM, Birkett C, Somers VK. Relationship between spectral components of cardiovascular variabilities and direct measures of muscle sympathetic nerve activity in humans. Circulation. 1997;95:1441–1448.
Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Heart rate variability: standards of measurements, physiological interpretation, and clinical use. Circulation. 1996;93:1043–1065.
Malik M, Camm AJ. Heart Rate Variability. Armonk, NY: Futura Publishing Co; 1995.
Brown TE, Beightol LA, Koh J, Eckberg DL. Important influence of respiration on human R-R interval power spectra is largely ignored. J Appl Physiol. 1993;75:2310–2317.
Piepoli M, Sleight P, Leuzzi S, Valle F, Spadacini G, Passino C, Johnston J, Bernardi L. Origin of respiratory sinus arrhythmia in conscious humans: an important role for the arterial carotid baroreceptors. Circulation. 1997;95:1813–1821.