The Shinken Database is a single hospital-based database that was established in June 2004 and includes data on all patients newly visiting the Cardiovascular Institute, Tokyo, Japan, excluding foreign travelers and patients with active cancer. Details of this database have been described elsewhere [12].

In the present study, a database of ECG results was used, which has been available since February 2010. From a total of 32,570 patients in the Shinken Database, 19,170 patients registered between February 2010 and March 2018 were extracted. After excluding patients with structural heart diseases (n = 4,915); patients aged < 20 years or > 90 years (< 20 or > 90 years; n = 168); and patients with index ECG showing indeterminate axis (R axis > 180°) (n = 76), pacing beats (n = 102), and atrial or ventricular tachyarrhythmia (n = 1,763), 12,837 patients were included in the present study. The structural heart diseases were defined as follows: valvular heart disease, moderate or severe stenosis or regurgitation on echocardiography; coronary artery disease; hypertrophic and dilated cardiomyopathy; and symptomatic heart failure [13].

The health status and incidences of cardiovascular events and all-cause death were obtained once per year from the medical records or the postal prognosis documents. In the present study, we included the follow-up data until March 2019 and excluded follow-up data from > 3 years after the initial visit to avoid an imbalance in the follow-up periods as a result of the different registration years (between 2010 and 2018) [13].

The 12-lead ECG was recorded by a GE ECG machine (GE CardioSoft V6.71 and MAC 5500 HD; GE Healthcare, Chicago, IL), and data were stored using the MUSE data management system [13]. Of the 639 parameters that had been automatically measured, 201 parameters (of which 9 were not lead-specific and 192 [16 × 12 leads] were lead-specific), including the relative coordinate points (i.e., the start point of P-wave), were excluded from the analysis [13]. Accordingly, the remaining 438 parameters (of which 6 were not lead-specific and 432 [36 × 12 leads] were lead-specific) were used in the analysis (Table 1).

We developed a predictive model for all-cause and cardiovascular (CV) death using 438 ECG parameters according to the following steps. Step 1: Univariable logistic regression analysis was performed for 438 ECG parameters (Fig. 1a, b; Wald statistics for each parameter are presented). Step 2: The Spearman’s coefficient of correlation was evaluated all combinations of the 438 parameters (438 × 437 = 191,406 combinations, excluding the pairing of A vs. A). The parameters combinations with correlation coefficients ≥ 0.9 (defined as “strong correlation”) were identified. Among them, those that demonstrated higher Wald statistics in Step 1 compared with any counterparts were selected for the next step. Parameters were also selected for the next step when they were not included in any pairs with “strong correlation”. Step 3: Among the ECG parameters selected in Step 2, parameters with statistical significance in the univariable logistic regression analysis (Wald statistics > 3.841458 [corresponding to P < 0.05] in Step 1) were selected for the final model. Step 4: Using the ECG parameters selected in Step 3, a prediction model was developed by a support vector machine (SVM) algorithm. For robust evaluation, a tenfold cross-validation method was employed, in which the study patients were divided into 10 similarly-sized groups according to the last digit of their study number (0 to 9), and the model was run 10 times using different combinations of training and testing datasets. For the first run, the testing dataset comprised the group with study numbers ending in 0, and the training dataset comprised the remaining nine groups; for the second run, the testing dataset comprised the group with study numbers ending in 1, and the training dataset comprised the remaining nine groups. The modelling was repeated like so for each of the 10 groups until the testing dataset comprised the group with study numbers ending in 9. The average values of permutation importance [14, 15] for each parameter and the average values of c-statistics were calculated, which evaluated the parameter importance and the model predictive ability, respectively.

For the final results, the following two values were described: (1) Relative importance of ECG parameters. The top 30 values of relative importance (%) were presented, which were calculated using the following equation: [average permutation importance of a parameter] / [average permutation importance of the top 1 parameter] × 100 (%), where the average permutation importance was the average of 10 permutation importance values obtained in the 10 training datasets. The full list of permutation importance values is presented in the Additional file 1: Tables S2 and S3. (2) Predictive capability: the c-statistics represented the predictive capability of the SVM models. C-statistics were separately evaluated for 10 training and 10 testing datasets, and their overall average values with standard deviation were described.

The characteristics of 12,837 patients are shown in Table 2. The patients in the study included 6,897 males (53.7%), and the mean age of all patients was 55.5 ± 15.0 years.

During the mean follow-up period of 320.4 days, all-cause deaths occurred in 55 patients (0.5 per 100 patient-years), among which 23 were cardiovascular deaths (0.2 per 100 patient-years).

Step 1: The Wald statistics of 438 ECG parameters in the univariable logistic regression analysis for all-cause death and CV death are shown in the order of ECG time phases (P, QRS, and ST-T) in Fig. 1a, b, respectively (gray and red bars; the full list is shown in Additional file 1: Table S1). Step 2: Spearman’s coefficient of correlation was evaluated for any pairs in the 438 ECG parameters. Among the 438 ECG parameters, 276 parameters did not have combinations with a correlation coefficient ≥ 0.9, thus, they did not have counterparts with a strong correlation. The remaining 162 had combinations with correlation coefficients of ≥ 0.9. Among them, we selected 54 and 52 parameters that had higher Wald statistics compared with any counterparts in Step 1 for all-cause and CV death, respectively. Accordingly, a total of 330 parameters (276 + 54) for all-cause death and 328 parameters (276 + 52) for CV death were selected for the next step. Step 3: Among the 330 and 328 parameters selected in Step 3, 109 and 70 parameters with statistical significance in the univariable models for all-cause and CV death, respectively, in Step 1 (Wald statistics > 3.841458) were selected for the final model (Fig. 1a, b, red bars). Step 4: Using the respective 109 and 70 parameters, the prediction models for all-cause and CV death were developed by an SVM algorithm. The results are shown below.

The major outcomes of the present study were as follows: (1) We developed a predictive model for all-cause death, in which the mean c-statistics of 10 model runs were 0.881 ± 0.027 for the training dataset and 0.927 ± 0.101 for the testing dataset. (2) The mean c-statistics of 10 model runs for CV death were 0.862 ± 0.029 for the training dataset and 0.897 ± 0.069 for the testing dataset. (3) ECG parameters with high permutation importance for both all-cause and CV death were concentrated in the QRS complex and ST-T segment.

Several studies have investigated the feasibility of using visible ECG parameters, including P wave characteristics [8, 16, 17], QRS morphologies [18, 19], ST-T segments, [9, 10] or QT duration, to predict mortality [20]. In a previous study, several ECG parameters, including heart rate > 75 bpm, QRS transition zone > V4, left ventricular hypertrophy, frontal QRS-T angle > 90°, prolonged QTc interval, and prolonged Tpeak-to-Tend interval, were identified as predictors of sudden cardiac death [6]. Our data were partially consistent with this previous study because the ECG parameters that potentially contributed to the prediction of all-cause and CV death were widely distributed among the P, QRS, and ST-T segment. However, when we comprehensively analyzed the ECG parameters contributing to the prediction of all-cause and CV death based on Wald statistics or parameter importance, the most important parameters were concentrated in the QRS complex and ST-T segment for both outcomes. Although the risk models in our study and the previous study [6] were similar to some extent, it is natural that the studies had different findings, as the former predicted all-cause and CV death in patients without structural heart diseases, whereas the latter predicted sudden cardiac death and included patients with structural heart diseases [6].

There have been many studies demonstrating the association between abnormal findings in the QRS complex and ST-T segment and CV mortality. Left ventricular hypertrophy [6] and fragmented QRS [19] have been demonstrated to be risk factors for CV death. The QRS complex is affected by electrophysiological impulse generation and propagation through the ventricles. [21] The pathology underlying QRS abnormality, such as ventricular fibrosis, inflammation, edema, fatty inflammation, ischemic cellular changes, or abnormal myocardial deposition of substances [21], can affect the current or future cardiac function, which may lead to a worse prognosis. The electrocardiographic strain pattern is associated with left ventricular concentric remodeling and scarring. Therefore, it is associated with the future development of various CV events, including heart failure or myocardial infarction, which result in increased mortality, even for patients free from CV diseases at baseline [9].

Our models demonstrated better predictive capability for all-cause death (the mean c-statistics of 10 models was 0.881 ± 0.027 for the training dataset and 0.927 ± 0.101 for the testing dataset) and CV death (0.862 ± 0.029 for the training model and 0.897 ± 0.069 for the testing model) than previous studies that used a few ECG parameters, in which the c-statistics for predicting death were 0.58 (maximal P wave duration) [22], 0.64 (minimal P′ amplitude in lead V1 and V2) [22], 0.61 (QRS area) [23], 0.55 (QRS morphology) [23], 0.51 (QRS duration) [23], 0.727 (QRS-T angle), [24] and 0.759 (model including clinical variables, such as age, sex, hypertension, diabetes, and ECG parameters) [6]. Not surprisingly, the high predictive capabilities of our models are due to the use of a large number of ECG parameters as consecutive values and the application of machine learning, which may sacrifice simplicity but prioritize the predictive capability [11].

In this study, we used hundreds of ECG parameters with automatic measurement. Such analysis would have sense when we excluded patients with heart diseases, because the characteristics in ECG with heart diseases are visually apparent. We intended to concentrate on the differences which are visually difficult to be distinguished. We thought small differences in numerical measurement of ECG parameters are affected by age, which should include the atherosclerotic changes in aorta, remodeling in heart, or simply the change of body shapes. For this purpose, we excluded patients with heart diseases, and consequently, the number of the endpoints in the present study (all-cause death or cardiovascular death) became very small.

Further, we should discuss the advantages and disadvantages of using machine learning approach in such prognostic studies. Initially, we intended to develop the clear and practical predictive model or risk score by Cox regression analysis. However, considering the small number of events, the multivariate analysis was clearly oversized in view of the number of events and raised the problem of statistical power. As we could not increase the statistical power (i.e., increase the incidence number of endpoints), we abandoned to use the statistical model like Cox regression analysis. Instead, we employed the machine learning algorithm which can work with a small number of events and a relatively large number of parameters.

Considering the possibility of overfitting by machine learning algorithm, we do not emphasize the differences in the effect of the individual ECG parameters on mortality. However, we believe our data provided a panoramic viewpoint and suggested that the ECG parameters affecting mortality were mostly concentrated in the QRS complex and the ST-T segment.