2.1 The study protocol and population
Healthy database (THEW identification: E-HOL-03-0202-003) with 202 24-hour ECG recordings and ESRD database (THEW identification: E-HOL-03-0202-016) contains 51 48-hour long-term ECG recordings were selected from the Telemetric and Holter ECG Warehouse (THEW, http://thew-project.org/datab ases.htm) at University of Rochester. Raw ECG data, as well as automated beat annotations reviewed and adjudicated manually, are available in both databases. All human data were obtained retrospectively from completed, Institutional-Review-Board-approved clinical research studies with subject de-identification. These trials complied with the Declaration of Helsinki and all subjects signed informed consent documents.
The ESRD database comprises Holter recordings with a sampling frequency of 1000 Hz from 51 ESRD patients with a high risk of death. All 51 ESRD patients underwent high-resolution 12-lead 48-hours continuous ECG monitoring, and THEW provides 12-lead raw ECG and heartbeat interval data. ESRD subjects receiving hemodialysis confirmed the history of diabetes or hypertension requiring treatment entered into this study. The ESRD patients were enrolled from February 13, 2009 to June 18, 2010, and they completed their 13-month follow-up evaluation. All patients Exclusion criteria included a history of chronic atrial fibrillation, with class I antiarrhythmic, pacemaker, implantable cardioverter-defibrillator device, cardiac resynchronization therapy device, female subject of childbearing potential not using medically prescribed contraceptive measures and subject unable to cooperate with the protocol due to dementia, psychological, or other related reason. The Healthy database comprises 24-h Holter recordings from 202 ostensibly healthy subjects. Subjects had (1) no overt cardiovascular disease or history of cardiovascular disorders; (2) no reported medications, history of high blood pressure and chronic illness, (3) a normal physical examination, (4) a 12-lead ECG showing sinus rhythm with normal waveforms (or a normal echocardiogram and normal ECG exercise testing in the presence of any questionable findings ECG changes). The ECG signals were recorded at a sampling frequency of 200 Hz. In order to reduce the influence of gender and age on ECG parameters, we try to match each ESRD patient with a healthy control with the same gender and close age. After excluding the poor signal quality and incomplete ECG recordings, 51 ostensibly age-matched healthy control subjects were selected from 202 healthy subjects and eventually enrolled in the present study.
2.2 ECG preprocessing
All long-term ECG recordings were analyzed with Kubios (Kubios 2.2，University of Eastern Finland，Kuopio), on which R waves were detected and labeled automatically. Heart-beat interval between 300 and 2000 ms, consecutive heart-beat interval differences ≤200 ms, and prolongations or shortenings ≤20% than the average of five preceding sinus rhythm heart-beat intervals were considered as sinus rhythm QRS complexes . Thereafter, automatic annotated results were carefully visual inspected and manually corrected by editing ectopic beats, arrhythmias and noise to suppress computational errors. Four-hour episodes of heart-beat intervals without naps and exercise within daytime (between 8 a.m. and 5 p.m.) were extracted from each recording for MSE and traditional HRV analysis [18, 19]. The 4-hour ECG recordings of ESRD patients used in the present study were selected after the ideal body weight to reduce the influence of volume overload on HRV. All ECG segments were selected from the same period to reduce the confounding effects that may occur due to sleep or diurnal rhythm. Furthermore, four-hour ECG segments of patients with ESRD after hemodialysis sessions while performing their usual daily activities were used for further analysis.
2.3 Traditional HRV analysis
Traditional techniques of HRV analysis are grouped into the time domain, frequency domain, and non-linear methods. The time-domain measures including mean heartbeat intervals (Mean RR), standard deviation of the heartbeat intervals (SDNN), square root of the mean of sum of squares of the differences between adjacent heartbeat intervals (RMSSD), and percentage of the absolute change in consecutive heartbeat interval exceeds 50 ms (pNN50) were calculated to represent the total variance and vagal modulation of heart rate . Based on the Fast Fourier transform spectrum, the frequency domain measures were computed from the power spectral density estimate for each frequency band including absolute power values of very low frequency (VLF, 0.0033–0.04 Hz), low frequency (LF, 0.04–0.15 Hz), high frequency (HF, 0.15–0.40 Hz) bands, total power (TP, 0.0033–0.40 Hz) and LF/HF power ratio . The VLF, LF, HF and TP were also transformed in natural logarithmic (ln) value. Five traditional non-linear measures were also taken into consideration to characterize the properties of HRV. SD1 denotes the short-term variability caused by respiration, whereas SD2 denotes the long-term variability with both calculated through the Poincaré plot method . Approximate entropy (ApEn) quantifies the single-scale complexity or regularity of the HRV time series by measuring the unpredictability of fluctuation patterns, and more uncorrelated random HRV signals usually produce higher ApEn value . As a technique for characterizing the nature of long-range correlations in time series, detrended fluctuation analysis (DFA) was applied in the present study to quantify slope α1 and α2 for characterizing the short-term and long-term fluctuations of HRV signal, respectively .
2.4 MSE analysis
The MSE technique was proposed to characterize complex structure of non-linear and non-stationary physiological signals in different temporal scales that ignored by traditional entropy methods. It comprises of two steps: 1) coarse-graining the time series in finite length into different time scales; 2) quantifying the degree of irregularity in each coarse-grained time series by sample entropy calculations [14, 15]. The quantified entropy values of coarse-grained time series then are represented as the function of time scale factors to evaluate the complex structure of physiological time series, and the features of the MSE curve can be extracted for clinical categorization in several diseases [16-19]. In-depth, details of this methodology have been previously described [14, 15]. In the present study, the complexity index (CI) of the HRV time series were quantified by curve fitting and calculating the area between the MSE curve and the axis of scale factors [18, 19]. The linear-fitted slope (Slope 5) and the area under the MSE curve between scales 1 and 5 (Area 1-5) were calculated to quantify the short-term complexity and to characterize the short-scale modulation pattern. Long time scale complexity was quantified by the fitted area under the MSE curve between scales 6 and 15 (Area 6-15) and between scales 6 and 20 (Area 6-20), respectively. Since low frequency drifts, high frequency non-stationarities and general hidded trends longer than 2 hours may lead to incorrectly increased irregularity and diminished sequence matching manifesting unpredictable effects on calculated sample entropy values. Empirical mode decomposition (EMD) is suitable for decomposition of non-stationary, non-linear physiologic time series and possesses advantages over wavelet and Fourier analysis because it employs a fully adaptive approach derived by means of a sifting process. In order to remove such effects, we used empirical mode decomposition (EMD) method for raw HRV time series filtering before performing MSE [18, 19].
2.5 Statistical analysis
Clinical data and parameters of ECG recordings were presented as median (25th and 75th percentiles). Gaussian distribution and homogeneity of variance tests were applied to determine the distribution and homoscedasticity of sample data. As a result of the non-normal distribution and heterogeneity of variance of some sample data, continuous variables were compared between different groups by the Mann-Whitney U test. For single predictive variable analysis using qualitative or categorical variables, Fisher's exact tests were applied for comparison between different groups. Correlations between clinical variables and independent factors that predicting all-cause death for ESRD patients were performed using Spearman’s correlation tests. The receiver operating characteristics (ROC) curve was created based on the sensitivity and specificity of HRV measures in predicting all-cause death in ESRD patients undergoing hemodialysis. The area under the ROC curve (AUC) gave an estimate of the overall discriminate ability (AUC=0.5 indicates no discrimination and an AUC=1.0 indicates a perfect diagnostic test). Statistical analyses were performed using SPSS version 20 software package (SPSS, Chicago, Ill, USA). The maximal hazards ratio and independent correlation of variables with mortality was determined by Cox regression analysis. Then, Kaplan-Meier event probability curves and log rank analysis of the dichotomized groups were obtained. For all statistical analysis, p values were corrected by the false discovery rate (FDR) method for multiple comparisons and p<0.05 was considered significant.