2.1. Study populations
Two datasets with 24-hour Holter recordings were used from the Telemetric Holter ECG Warehouse (THEW) of the Center for Quantitative Electrocardiography and Cardiac Safety at University of Rochester Medical Center (Rochester, NY, USA) [5
]. The healthy database (E-HOL-03-201-003) included of 154 subjects (37 ±15 yrs), 79 males (37 ±14 yrs) and 75 females (37 ±16 yrs).
The AMI database (E-HOL-03-160-001) was divided into two groups based on infarct location. Subjects with transient ST depressions were excluded. Group A consists of subjects with a myocardial infarct with anterior or anterior lateral location, and includes 25 subjects (57 ±16 yrs), 18 males (51 ±13 yrs) and 7 females (71 ±16 yrs). Myocardial infarct locations for group B were inferior or inferior lateral. Group B includes 28 subjects (56 ±13 yrs), 22 males (55 ±12 yrs) and 6 (62 ±15 yrs). The two groups contained recordings for both the acute and stable phase of the myocardial infarct. Further clinical characteristics of these populations can be found on the THEW project website (http://www.thew-project.org
2.2. ECG preprocessing
The ECG recordings were digitally recorded at 200Hz. The ECGs were filtered using a fifth order IIR forward-backwards bandpass filter with a lower cutoff frequency of 0.5 Hz and a higher cutoff frequency of 35 Hz to remove the effects of baseline wander and muscle noise. The lead configuration used in these two datasets (Healthy and AMI) was similar and included 3 pseudo-orthogonal leads (X, Y and Z). In this study we only used lead X. All ECGs were recorded with the same Holter equipment and exported from the THEW file server in ISHNE format for both the raw ECG signals and the corresponding cardiac beat annotation.
2.3. Rpeak detection and beat selection
The Holter recordings from the THEW provided location of the center of gravity (COG) of the QRS complexes. In this study, we used the Rpeak as a reference for the detection of the onset and the end of cardiac beats. The location of the COG was used as an aid to detect these fiducial points. Using a 150ms search window centered on COG, the largest maxima in the window was defined as Rpeak.
Non-sinus cardiac beats were identified using the annotation information from the THEW, and these beats were excluded from the analysis. Furthermore, RR intervals and cardiac beats prior and after the non-sinus cardiac beats were excluded.
The cardiac beats associated with HRs below 60 BPM or above 100 BPM were discarded from our analysis in order to have a sufficient number of beats to compute a representative cardiac beat for a given HR.
Finally, only cardiac beats from the diurnal period were included in order to consider similar autonomic regulation of the heart. The diurnal period was defined between 7:00 am and 11:00 pm.
2.4. RR-bin and ECG profile definitions
The width of each bin was defined as 1/fs, where fs is the sampling frequency of the ECG signal. Bins with less than 20 cardiac beats were considered not representative.
The cardiac beats associated with a given RR bin (previous RR interval was used) were aligned using the Rpeak. Then for each collection of beats in a bin we calculated a ”median beat”. The ”median beat” was computed from the median values of the amplitude of all synchronized samples across beats in a given RR-bin. We used the median instead of the mean to ensure a better estimate in case of the presence of noisy beats or undetected artifacts.
Subject profiles were used to calculate population profiles. For the subject-based profiles, ”median beats” were calculated for all RR-bins. For the population-based profiles, the ”median beats” from subjects belonging to the same study population, and having the same previous RR interval were grouped and normalized using the amplitude of the Rpeak. For each population-based RR-bin, the reported ”median beat” was calculated by synchronizing the subject-specific median beats on the Rpeak.
2.5. Detection of Tpeak
To detect the location of Tpeak in the profiles, we used a wavelet-based technique [6
]. This delineation method was applied on the ”median beats” in all bins to get the amplitude of Tpeak for that specific RR-bin.
Filtering and beat extraction were implemented in C++ using BOOST libraries. All graphs of the profiles and statistics were done using MATLAB 7.9.0 (Mathworks Inc, Natick, MA, USA).