We performed a prospective community-based cohort study in Taiwan, the HALST study (Supplementary Material), a community-based cohort study, that enrolled 5,380 study participants (12–14). Participants were enrolled from December 2008 through March 2013, and were followed up until April 2019. We collected three serial 12-lead ECGs and other relevant clinical and demographic information prospectively at the time of enrollment according to the study protocol. All analysis were conducted between May 2019 and September 2020.

We excluded individuals with a history of any cancer, significant heart diseases (e.g., documented coronary artery disease, severe heart failure (New York Heart Association class III/IV), myocardial infarction, severe valvular diseases, any kinds of cardiomyopathies and infectious cardiac diseases), ventricular conduction delay (2nd or 3rd degree atrioventricular block), atrial fibrillation or flutter, congenital short or long QT syndrome (15, 16), Brugada-type ECG (17), catecholaminergic polymorphic ventricular tachycardia (18), ventricular pre-excitation, and implanted pacemakers. The final study cohort population consisted of 4,615 individuals, constituting a random sample representative of the entire source population. All those who died of cardiac or other causes (n = 87 and n = 683, respectively) during the follow-up (mean follow-up time 95.1 ± 21.9 months) were identified.

The clinical outcomes that are considered in this study are all-cause death, cardiovascular death, and unexplained death. Death from cardiovascular causes was defined by the ICD codes I01-I02.0, I05-I09, I20-I25, I27, and I30-I52. The definition of unexplained deaths are sudden deaths with unexplained causes, occurring in an individual older than 1 year, according to 2020 AHPRS/HRS expert consensus statement (19). Taiwan maintains extensive administrative registers that record every death in the country. Two clinicians blinded to the ECG results determined the causes of death by examining death certificates from the National Taiwan Institute of Health and Welfare. The causes of death were adjudicated by a committee of experienced cardiologists, who were blinded to the data from the electrocardiographic analyses.

The 12-lead ECGs (Hewlett Packard, USA) were recorded at the standard settings of 10 mm/mV and 25 mm/s. PR and QRS intervals were computed automatically, while QTc was computed using Bazett's formula. We selected ECG variables, that are associated with mortality and relatively easily obtained in clinical practice, for risk analysis. The criteria for abnormality for each of the ECG parameters were defined as follows: PR interval >220 ms, QRS duration >110 ms, spatial QRS-T angle >90°, corrected QT interval (QTc) >450 ms in men or >460 ms in women, heart rate > 80/min, ST segment elevation in lead aVR, early repolarization pattern (ERP) in inferior or lateral leads (lead II, III, aVf; V4–6), fragmented QRS (fQRS) (20, 21) and left ventricular hypertrophy (LVH) by Sokolow-Lyon criteria or modified Cornell Criteria (22, 23). Each abnormal ECG variable counted for 1 point.

ECGs were displayed on a 24-in computer screen in multiple formats, enabling careful classification of slurring on the downslope of the R and J waves. All ECGs were analyzed and interpreted independently in random order by two trained cardiologists, who were blinded to clinical data and follow-up status. In case of divergent results, a third blinded cardiologist re-interpreted the ECG, and a preliminary decision on each ECG parameters was achieved by majority vote. After the preliminary decision on each ECG parameters, two trained cardiologists jointly re-assessed, and a final decision was reached by consensus. The inter-observer agreement was k = 0.97, and agreement proportion = 0.98.

From a total of 5,380 relatively healthy and ambulatory individuals, enrolled in this study (Supplementary Figure 1), 755 individuals were excluded due to cancer, underlying severe cardiovascular diseases (e.g., myocardial infarction or pacemaker implantation), or missing follow-up information. Finally, 4,530 individuals with complete data were used for analysis (Supplementary Table 1).

To evaluate whether ECG parameters were associated with the death outcomes, we performed a log-rank test using both univariate and multivariate approaches. As shown in Table 1, LVH, QTc and the PR interval were significantly associated with the primary outcome, all-cause death (P < 0.05). The hazard ratios (HRs) of LVH, QTc and the PR interval were 1.39 [95% CI = (1.16–1.67)], 1.31 [95% CI = (1.07–1.61)] and 1.40 [95% CI = (1.01–1.95)], respectively, in the multivariate model. Similarly, the LVH was a significant predictor of CV death [HR = 2.37, 95% CI = (1.48–3.79), P < 0.05] (Supplementary Table 2), and PR interval was the risk predictor of unexplained death [HR = 2.63, CI= (1.24– 5.57), P < 0.05] (Supplementary Table 3).

To further evaluate whether ECG parameters can be used to stratify people into different risk groups, we divided subjects into subgroups based on how many ECG abnormal parameters they displayed. The baseline characteristics of each ECG group are summarized in Table 2. Comparison among all groups were conducted to report p-values (Table 2 and Supplementary Table 4).Some clinical factors showed significant differences among the 4 groups, including age, systolic and diastolic blood pressures, smoking and hypertension (P < 0.05).

To explore whether the EA score was associated with the survival outcomes, we examined the proportion of death events for each EA level. The results are illustrated in Supplementary Figure 2; the proportion of events for EA subgroups increased with each additional ECG abnormal parameter. Group with 3 ECG abnormalities exhibited the highest proportion of events for outcomes, all-cause death (10.25, 13.04, 18.39, and 19.82%), CV death (0.78, 129, 2.16, and 2.80%) and unexplained death (0.78, 1.29, 2.16, and 2.80%), respectively. We further performed a secondary analysis, using only two EA groups: low and high. The samples with no or one ECG abnormal parameter (0 or 1) were classified as low EA, whereas samples displaying two or more ECG abnormal parameters were classified as high EA. As shown in Figure 1, the high EA group showed a higher proportion of events for all-cause deaths (18.83%), CV deaths (2.82%) and unexplained deaths (2.36%) compared to that of the low EA group (11.97%, 1.13%, and 1.10%, respectively).

Furthermore, a competing risk survival analysis of the two-level EA groups were performed by using the Gray's tests to elucidate their associations with each of the death outcomes studied, while taking into consideration the possibility of the event to have an affect due to deaths by other causes (Figure 2). Notably, the high EA group showed significantly higher cumulative incidence in all three outcomes: all-cause death (P = 0.004, Figure 2A), CV death (P < 0.001, Figure 2B), and unexplained death (P = 0.006, Figure 2C). The detailed classification (using four EA categories) of each group and its association with the different death outcomes is shown in Supplementary Figure 3, which showed similar pattern as that of the two-level EA groups for all three events. In conclusion, these results indicate that ECG parameters can help in identifying people at a higher risk for all-cause-death, cardiovascular death, or unexplained death.

Supplementary Tables 5–13 displays the results for the scaled Schoenfeld residuals testing that were conducted for all three study outcomes to check for the validity of the proportional hazard assumption for the covariates under consideration, in each of the proposed prognostic models in this study. The findings along with the global tests were not found to be statistically significant, implying that the proportional hazard assumption is valid for all models. Cox - proportion hazards regression was performed on an age and sex adjusted univariate model and a multivariate adjusted prognostic model, to check if the EA score is a significant predictor for all three endpoints all-cause death, CV death and unexpected death. The results for the primary outcome, all-cause death are summarized in Table 3. Notably, the EA (high, low) was a significant predictor in the age- and sex-adjusted model (HR: 1.31; 95% CI= (1.12–1.53), P = 0.0008) and the multivariate model (HR: 1.28; 95% CI = (1.09–1.50); P = 0.002) after considering all traditional risk factors. The c-index was 0.786 in the multivariate model, suggesting a good-fit of the prediction model and the real death events in all-cause death. As shown in Supplementary Tables 14, 15, the EA score was a significant predictor for both CV death, and unexplained death, respectively.

To evaluate the performance of the multivariate adjusted prediction model in comparison with the traditional model, discrimination analysis was conducted using a 10CV design. Table 4, Supplementary Table 16 lists the average and the standard deviations of the C- indices over all 10CVs for both the models and for all three study outcomes. Harrel's c-index values for all three outcomes were reported to be 0.786, 0.798, and 0.851, implying that the ECG model is an effective predictor of the study events and can correctly discriminate patient's survival for a given patient pair. Furthermore, comparison of the c-indices between ECG model and traditional model, using a bootstrapping strategy, revealed that adding ECG parameters can significantly improve the predictive capability (P < 0.001), based on our data.

Finally, a calibration analysis was conducted to examine if the predicted events were significantly different from the observed events, for all three study outcomes. Figure 3 displays the results of the calibration analysis that was performed using a 10CV for each of the 10 years where the average difference of the proportions between the predicted and observed events were plotted for both ECG model (high, low) and traditional model. For all-cause death the differences in the proportions were ≤5% for the first 5 years, while for both CV death and unexpected death the differences in the proportions were negligible (<3%) for all 10 years. The higher differences that were observed for all-cause death after 5 years may be attributed to the fact that these elder patients may die of diseases other than cardiovascular ones, such as cancers. Similar patterns were observed for the other EA score model (0, 1, 2, and 3) (Supplementary Figure 4) and traditional models (Figure 3 and Supplementary Figure 4). Supplementary Tables 17–22 further lists the detailed calibration analysis results for both ECG model and traditional model, for all three outcomes. All results suggest that our prediction model can help to identify subjects with a high risk of a death event among the elderly.