Ej Bowers

Principal Component Analysis as a Tool for Analyzing Beat-to-Beat Changes in ECG Features: Application to ECG-Derived Respiration

An algorithm for analyzing changes in ECG morphology based on principal component analysis (PCA) is presented and applied to the derivation of surrogate respiratory signals from single-lead ECGs. The respiratory-induced variability of ECG features, P waves, QRS complexes, and T waves are described by the PCA. We assessed which ECG features and which principal components yielded the best surrogate for the respiratory signal. Twenty subjects performed controlled breathing for 180 s at 4, 6, 8, 10, 12, and 14 breaths per minute and normal breathing. ECG and breathing signals were recorded. Respiration was derived from the ECG by three algorithms: the PCA-based algorithm and two established algorithms, based on RR intervals and QRS amplitudes. ECG-derived respiration was compared to the recorded breathing signal by magnitude squared coherence and cross-correlation. The top ranking algorithm for both coherence and correlation was the PCA algorithm applied to QRS complexes. Coherence and correlation were significantly larger for this algorithm than the RR algorithm(p < 0.05 and p < 0.0001, respectively) but were not significantly different from the amplitude algorithm. PCA provides a novel algorithm for analysis of both respiratory and nonrespiratory related beat-to-beat changes in different ECG features.

Respiratory rate derived from principal component analysis of single lead electrocardiogram

We used principal component analysis to derive the respiratory rate from single lead ECGs. In this algorithm the respiratory induced beat-to-beat variability of the ECG is described by the coefficients of the principal components. Subjects were asked to breathe at different rates and naturally while respiration and ECG were recorded. Breathing rate, determined by Fourier analysis, was compared for the ECG derived respiration obtained by principal component analysis and the recorded respiratory signal. Across the different breathing patterns the mean absolute differences between reference respiratory rate and ECG respiratory rate were 0.2 breaths per minute (bpm) or less. In all cases the respiratory rate was accurately determined from respiratory surrogates derived from principal component analysis of the single lead ECGs.

An algorithm to distinguish ischaemic and non-ischaemic ST changes in the Holter ECG

Changes in the ECG ST segment are often observed in patients with myocardial ischaemia. However, non-ischaemic changes in ST level are also common thereby limiting ischaemia detection accuracy. The aim of this study was to devise an algorithm and determine its accuracy in distinguishing between ischaemic and non-ischaemic changes in the ECG ST-segment, using expertly annotated ECG data sets as a gold standard reference. The algorithm considered only the change in ST relative to a baseline ST level (/spl Delta/ST) provided by the PhysioNet database, and based on simple level thresholding within specified time windows. An initial score of 82.3% (accuracy 91.1%, with sensitivity 99.0% and specificity 88.8%) was achieved for the learning set. By making slight modifications to the algorithm and introducing principal components of ST it was not possible to improve the original algorithm. The original algorithm was therefore left as our challenge entry achieving an accuracy of 90.7% for the test data set (score of 81.4%, entry 1, 1 May 2003).

Interaction between cardiac beat-to-beat interval changes and systolic blood pressure changes

Abstract.This study assessed the interaction between cardiac beat-to-beat interval changes and systolic pressure (SP) changes. Twenty subjects breathed regularly following displayed breathing signals at 4, 6, 8, 10, 12 and 14 breaths per minute, each for 5 minutes. ECG, non-invasive blood pressure (Finapres) and respiration waveforms were recorded. Time offsets between the cyclic patterns of RR-interval and SP changes were calculated. Displayed breathing signals were well followed; the mean correlation between displayed and recorded breathing signals ranged from 0.72 to 0.86. The time offset between RR-interval peaks and subsequent SP troughs decreased with increasing respiration rate, 3.8 ± 1.7 s, 3.5 ± 0.7 s, 3.1 ± 0.6 s, 2.6 ± 0.4 s, 2.3 ± 0.4 s and 2.0 ± 0.4 s mean ± SD at 4, 6, 8, 10, 12 and 14 breaths per minute respectively. The relationship between mean time offset and frequency was significant (p < 0.001), with a 95% prediction interval of ± 0.24 s. Published data showed no relationship between time offset and frequency, with a 95% prediction interval of ± 2.8 s. However, when the offset definition proposed in this research was applied to these data, a significant relationship (p < 0.01) was evident, with a 95% prediction interval of ± 1.5 s. In conclusion, apparently contradictory previous findings achieve good consensus when a standardized method for presenting results is applied. A delay exists between RR-interval and blood pressure changes, and this delay varies with breathing frequency.

Can paroxysmal atrial fibrillation be predicted

Atrial fibrillation is an ECG rhythm with a significant mortality due to stroke. The objective of this study was to detect those patients most likely to develop atrial fibrillation, and to identify ECGs closest to the onset of fibrillation. Our hypothesis was that patients with atrial fibrillation would have atrial ectopy, and the frequency of this activity would increase prior to onset of fibrillation. From a learning set of 100 30-minute ECGs from 50 patients, 25 without atrial fibrillation (normal) and 25 who subsequently developed atrial fibrillation, an algorithm was developed to detect the presence of ectopic beats using R-R interval data. In the learning set, 37/50 abnormal and 34/50 normal patients were identified, giving a potential screening accuracy of 71%. As a prediction test to detect the ECGs closest to atrial fibrillation, 19/25 were correctly identified. For the test set, a total of 29/50 were correctly assigned to the normal and fibrillation groups, and a 39/50 score obtained in predicting the onset of atrial fibrillation.

Effects on baroreflex sensitivity measurements when different protocols are used to induce regular changes in beat-to-beat intervals and systolic pressure.

Baroreflex sensitivity is becoming an important clinical measurement. Nevertheless there is no recommend standard measurement protocol. This study assessed the ability of eight protocols to induce regular changes in cardiac beat-to-beat interval and systolic pressure (SP), and the effect each protocol had on baroreflex sensitivity (BRS). Twelve subjects had changes in cardiac beat-to-beat intervals and SP levels induced at 8 times a minute by following 8 different protocols, each for 3 min. These comprised breathing in a supine and standing posture, breathing through a resistance, breathing into a closed orifice (the breathing protocols), and performing handgrip exercises, being rocked, having legs raised and lowered, and being presented with mental arithmetic questions (the non-breathing protocols). Induction success of each protocol was determined by the percentage of cardiac beat-to-beat interval and SP level signals with a peak at 8 times per minute in their frequency spectra. The consistency of the induced changes was measured by a signal-to-noise ratio (SNR). BRS was calculated from the frequency spectra. The induction success was 85% for breathing and 31% for non-breathing protocols. The consistency of cardiac beat-to-beat interval changes was highest with supine breathing (SNR = 1.6 +/- 0.3) and resistance breathing (SNR = 1.5 +/- 0.5) protocols. The consistency of SP level changes was highest with resistance breathing (SNR = 1.0 +/- 0.3) and breathing into a closed orifice (SNR = 1.0 +/- 0.5) protocols. BRS values in the supine breathing protocol (24 +/- 10 ms mmHg(-1)) and the handgrip protocol (32 +/- 3 ms mmHg(-1)) were significantly greater (p < 0.05) than for standing breathing (11 +/- 5 ms mmHg(-1)), resistance breathing (17 +/- 8 ms mmHg(-1)) or breathing into a closed orifice (12 +/- 5 ms mmHg(-1)) protocols. Different protocols have different induction successes and degrees of effectiveness in inducing cardiac beat-to-beat and SP level changes. BRS is affected by the induction protocol used, highlighting the need for a standard measurement protocol.

Heart rate variability characteristics required for simulation of interval sequences

Fifty sequences of PhysioNet R-to-R interval data, covering periods of between 20 and 24 hours, were classified into real or simulated groups. The RR interval characteristics were investigated in both the time domain and frequency domain. Eleven characteristics were analysed, and the range of measurements for each was studied for outliers from the main distribution. In the time domain, a restricted pattern of RR interval distributions classified 4 sequences as abnormal, and a reduced RR variability produced 18 classifications, with an overlap of 8, giving a total of 14/50 as abnormal. In the frequency domain, abnormally restricted very low frequency characteristics produced 26 classifications as abnormal with 10 overlaps giving a total of 16. The low frequency to high frequency ratio classified 4 as abnormal, but three of these were already detected by abnormal low frequency characteristics, giving a total of 17 classified in the frequency domain. Of the 17 classified in the frequency domain and of the 14 in the time domain there was an overlap of 9, resulting in 22 abnormal classifications, and suggesting that these were simulated. When PhysioNet assessed this classification a correct grouping of 100% was achieved on a single entry (reference 20020426.082234).

Simulation of cardiac RR interval time series

An RR interval simulator was developed as part of the Computers in Cardiology Challenge (entry no. 184). The simulator was based on observed physiological changes in normal subjects. A template, which represented slow trends in RR interval during 24 hours, was derived from a number of parameters that represent sleep and wake states. The variations about the template were generated so that the frequency spectrum of the final signal was similar to data from normal subjects. The frequency spectrum of normal RR intervals shows a strong 1/f component, a peak at around 0.1 Hz and a peak corresponding to respiratory rate between 0.15 Hz and 0.4 Hz. These were simulated by adding to the template pink noise, a number of random phased sinusoids at around 0.1 Hz and a signal that represented respiration. Variations due to respiration were dependent on sleep and wake states.

The ebb and flow of heart rate variability: simulation of 24 hour heart rate time series using time series data from naturally occurring phenomena

Current RR time series simulations are distinguishable from real data by automatic algorithms. We hypothesised that RR time series simulations could be improved by using time series data from naturally occurring phenomena. 20 records of annual river flow data for the river Tyne in north eastern England were obtained. Each river flow data record was used to generate a single 24 h simulated RR time series with the property of self similarity. We compared the standard frequency parameters ULF, VLF, LF and HF normalised to the total power, for the simulated RR, with those from physiological data from 20 subjects. The river flow data produced realistic simulations of RR time series with significant differences between physiological and simulated series for VLF only. Time series data from river flow or other naturally occurring phenomena may provide useful components in producing RR time series with more realistic characteristics than current artificially generated data

Analysis of RR interval and fibrillation frequency and amplitude for predicting spontaneous termination of atrial fibrillation

We assessed characteristics of atrial and ventricular activity from the ECG for predicting the oflset of atrial fibrillation for the 2004 PhysioNet/Computers in Cardiology Challenge. Seven parameters were analysed with five based on the statistical characteristics of the RR interval (mean, standard deviation, skewness, kurtosis and median beat-to-beat change), and fibrillation frequency and atrial signal amplitude. The power of the parameters to predict termination of the arrhythmia was assessed individually and in combination using linear discriminant analysis ( D A ) and a?t