For this study, 226 patients aged from 18 to 80 years who were admitted for AMI to the Cardiology Department of the 1st hospital of the Harbin Medical University in Harbin, China from January 2009 to December 2009 were enrolled. AMI was diagnosed at the time of admission by at least 2 of the following findings: chest pain for ≥20 min; creatine kinase-MB more than double the upper normal limit of our laboratory; ST-segment elevation ≥0.1 mv in at least two limb leads or ≥0.2 mv in at least two contiguous precordial leads; and primary motion abnormality of the left ventricular regional wall in the echocardiography or MRI examination. We excluded patients if they: (1) had severe valvular, pulmonary, hepatic, renal or other severe concomitant noncardiovascular disease; (2) were logistically unable to participate; (3) exhibited sinus rhythm for <80% of the total 24-hour Holter recording; (4) had atrial flutter, atrial fibrillation or pacemaker rhythm; (5) had inadequately analyzable electrocardiographic data; and/or (6) had inadequate follow-up data. Furthermore, all of patients underwent routine cardiac medication and AMI treatment according to the contemporary guidelines [8].

The analysis of electrocardiographic data and the monitoring of patients were approved by the Institutional Ethics Committee for human research of the Harbin Medical University. Since the data collection was noninvasive and did not affect routine clinical management of the patients, the Ethics Committee deemed oral informed consent to be sufficient for patient participation.

Holter electrocardiograms were recorded for 24 hours during the second week after infarction (average 8–14 days) for patients that were stable during this period. The recordings were digitized at 128 Hz, automatically processed with a GE Holter system (GE medical, USA), and scanned with GE Holter analysis software following standard procedures as specified by the manufacturer. Manual verification of uncorrected QRS classifications and reduction of noise was conducted as needed. Meanwhile, we assessed LVEF by single-plane echocardiography one week after index infarction, with <0.35 predefined as seriously abnormal. Other risk predictors included age, history of myocardial infarction, diabetes mellitus, hypertension, and medical treatments, which were determined by the physician using a standardized protocol.

Machine learning algorithms require data to be represented by features, such as the indices that occur in an electrocardiographic recording. In this study, a total of 10 indices were extracted from every Holter recording. The electrocardiographic features of the patients are shown in Table 1 and include the following: indices of HRV; heart-rate turbulence; mean heart beat; DC; and acceleration capacity of heart rate. As proposed by the Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology [9], we calculated the following conventional risk predictors from time domain measures of HRV in the 24-hour Holter recordings. Time domain measures of the R-R time series were used to quantify the variability of the R-R intervals, including standard deviation of all normal-to-normal intervals [SDNN (ms)], average standard deviation of all normal-to-normal intervals (ms), standard deviation of average normal-to-normal intervals (ms), mean normal-to-normal intervals (ms), and triangle index of geometric figure parameter of the R-R time series. The spectral measures of HRV were analyzed using the methods recommended by the Task Force, and the cut-off values of reduced spectral indexes were also predefined. These traditional indices of HRV have a proven ability to predict mortality [10] and are easy to calculate. High resting heart rate has been associated with cardiac death [11]. In this study, mean heart rate was calculated as heart beat number per hour. We assessed other newer risk predictors of heart-rate dynamics as well. Heart-rate turbulence indices, including turbulence onset and turbulence slope (ms), reflect the baroreflex mediated short-term oscillation of cardiac cycle lengths after spontaneous ventricular premature complexes. These were calculated in the instances when ventricular premature beats were found in the Holter recordings [12]. Turbulence slope was defined as the maximum slope of the regression line assessed over any sequence within five subsequent sinus R-R intervals during the first fifteen sinus heart beats following an ectopic ventricular beat. A decreased heart rate turbulence is related to autonomic dysfunction and has a powerful ability to predict sudden cardiac death [13]. Two newer Holter-based risk variables, which are developed by a phase rectified signal averaging technique—namely, DC (ms) and acceleration capacity (ms) of heart rate—were calculated to provide a separate assessment of autonomic modulation of the heart rate. DC reflects the regulation of the vagus nerve on heart rate. This is of clinical importance because vagus withdraw has been regarded as an important mechanism of sudden death [14]. Furthermore, these features are all thought to reflect functional changes of the autonomic nervous system.

In machine learning, SVM are supervised learning models with associated learning algorithms that analyze data and recognize patterns. SVMs are used for classification and regression analyses, and the classification strongly depends on the available feature set and the tuning of hyper-parameters [15]. We used SVMs to analyze the predictive power of the combined indices from every Holter recording. In this study, SVM was developed to maximize the margin of the hyper-plane dividing electrocardiographic data into two classes, survivals and cardiac deaths, by finding the best hyper-plane that separates clusters of features represented in an n-dimensional space. A hyper-plane can be written as the set of points X satisfying W^TX + b = 0, where W^T is a normal vector perpendicular to the hyper-plane, X is the vector of electrocardiographic features, and b is bias or offset of the hyper-plane from the origin. Inputs of SVM are mapped onto a high dimensional feature space via kernel functions, and the optimal hyper-planes are constructed to separate samples into two classes [16]. SVMs were trained and tested using the leave-one-out and cross-validation methods, which were applied to evaluate the accuracy of classification. At every step, one electrocardiographic data subset was left out from the total training dataset and treated as undefined data. The classification model was constructed with the remaining data and the algorithm classified the “left-out” electrocardiographic data. The same procedure was applied to the training dataset. The results of SVM were normalized in the range from 0 to 1. We prospectively defined a cut-off point as >0.5 indicated higher probabilities of death, while <0.5 indicated higher probability of survival. Through training and testing, we finished feature selection and searched for the optimal vector that had the greatest predictive power.

At the same time, comparisons were made with LVEF, SDNN and DC using the identical dataset. The SDNN measured from the 24-hour R-R intervals was chosen as a conventional index of HRV, with a value <70 ms predefined as abnormal [10]. Meanwhile, DC measured from the 24-hour recording was chosen as a novel index of heart-rate dynamics, with a value <4.5 ms predefined as abnormal [14].

Patients with AMI underwent physical examination 6 months after discharge and annually thereafter by telephone assessment. The patients were followed for a mean of 28 months after AMI. In cases of death, the causes were verified from the hospital records, death certificates, autopsy records and telephone interviews with either the primary physicians or those who had witnessed the death. An independent endpoint committee determined the cause of death according to the available data from the aforementioned sources. Witnessed deaths due to cardiac diseases including myocardial infarction, heart failure, severe malignant arrhythmia and those occurring after attempted resuscitation by defibrillator were all classified as cardiac death. Unwitnessed deaths occurring in a previously asymptomatic patient with no other life-threatening diseases was classified as cardiac death as well. Based on these determinations, the electrocardiographic data were then classified as survivals or cardiac deaths. To avoid bias, the computer center received only electrocardiographic data and classification results without other clinical information of patients.

We calculated receiver-operator characteristic curves to evaluate prediction accuracy of SVM models, LVEF, and conventional risk indices of HRV and DC. We quantified receiver-operator characteristic curves by taking the integrals of the curves (area under the curve; AUC), with the continuous variable being set to a plurality of different critical values, to calculate a series of sensitivity and specificity thresholds. The sensitivity was used as the ordinate and negative positive rate (1- specific) as the abscissa to plot the curves. The AUC of a method was prospectively defined as the statistical measure of the predictive power. The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of each variable were taken into account [17]. The difference between two receiver-operator characteristic curves was compared by non-parametric test, with a P value less than 0.05 considered statistically significant.