The original study population consisted of 315 patients with chronic CC enrolled in the protocol-based clinical follow-up program of patients from the CC outpatient clinic of the Clementino Fraga Filho University Hospital of the Federal University of Rio de Janeiro (HUCFF-UFRJ) between 1992 and 2017.10 Patients were excluded if they showed malignancy or severe noncardiac comorbidities (including psychiatric disorders, advanced liver disease, or renal disease), if they were lost to follow-up, or had prior cardiac resynchronization therapy, resulting in a screened population of 220 participants (figure 1).

Figure 1.

Flowchart of the study population. AB, ablation surgery; ICD, implantable cardioverter-defibrillator; PM, pacemaker; SCA, sudden cardiac arrest.

(0.14MB).

In each patient, a single 24-hour ECG Holter recording (sampling frequency 128Hz) was selected. When multiple recordings were available (52 patients, median inter-recording interval, 3.5 [4.9] years), the most recent recording before the primary endpoint was considered.

The primary endpoint included SCA, appropriate implantable cardioverter-defibrillator therapies or ablation procedures, as these interventions address similar electrophysiological and arrhythmogenic mechanisms and may represent prevented SCA events.11,12 SCA was defined as sudden cessation of cardiac activity with unresponsiveness, absence of normal breathing, and absence of signs of circulation.13 All CC patients who experienced SCA were referred for interventions (implantable cardioverter-defibrillator therapies or ablation procedures) within the framework of the Brazilian Unified Health System.13 Patients were excluded if they had a prior pacemaker, implantable cardioverter-defibrillator or ablation procedures before ECG-Holter monitoring, had ECG analog-to-digital conversion anomalies, or were classified as high risk by the RS (figure 1).

All patients received treatment that followed protocols previously established by the National Health Surveillance Secretariat/Ministry of Health for Chagas heart disease in Brazil. This study was approved by the institutional ethics committee of the Federal University of Rio de Janeiro (CEP-HUCFF-UFRJ) and written informed consent was obtained from all participants.10 Detailed information on the clinical setting and patient management can be found in the methods section of the supplementary data.

The RS3 was calculated for each patient at each medical evaluation (1 week before ECG-Holter acquisition) to estimate the risk of mortality. This score (0-20 points) is based on New York Heart Association functional class III or IV (5 points), cardiomegaly on chest radiography (5 points), left ventricular systolic dysfunction on echocardiography (3 points), nonsustained ventricular tachycardia on 24-hour Holter monitoring (3 points), low QRS voltage (2 points), and male sex (2 points). Patients were stratified into low- (0-6 points), intermediate- (7-11 points), and high-risk (12-20 points) categories.

Quantification of the 4 ventricular restitution indices was performed on each ECG recording following previously described methodology.14 First, a standard ECG preprocessing was applied (see the methods of the supplementary data). Then, an RR interval (RR) histogram was derived from each individual 24-hour recording using 10-msec-wide bins (figure 2A) and only those RR bins with more than 75 occurrences were included. This approach favors highly recurrent RR intervals and minimizes the influence of rhythm irregularities. Next, the 2 most distant RR bins from the median (RR1 and RR2) (figure 2A, green, red, and blue bins, respectively), distributed symmetrically around this median, were chosen as those defining the maximum intraparticipant RR range, ΔRR, ie, ΔRR=RR2-RR1. Then, the beats associated with the RR interval within these 2 bins were used for the computation of the 4 ECG-based indices, which can be divided into 2 groups:

Figure 2.

Quantification of the ECG indices of ventricular restitution. A: RR histogram, showing the median RR interval (green) and the maximum intraparticipant range, ΔRRI (red and blue). B: quantification of the temporal ECG indices of ventricular restitution. C: quantification of the morphological ECG indices of ventricular restitution.

(0.41MB).

Temporal ECG indices of ventricular restitution

The indices ΔαQT and ΔαTpe were defined as:

where QTmRR1 and QTmRR2, and TpemRR1 and TpemRR2 measure the median duration of the QT and Tpe intervals, respectively, of the beats measured at bins with RR intervals of RRI1 or RRI2 (figure 2B).

Morphological ECG indices of ventricular restitution

Two mean warped T-waves, representing the average T-wave morphology at RR1 and RR2, respectively, were computed using a previously described warping method15 (figure 2C). To quantify the morphological differences between the 2 T-waves, we used the index dw, which quantifies the level of warping required to align any 2 T-waves15:

where the warping function γ*(tr) optimally relates the temporal sample values of each T-wave (figure 2C, center).

The T-wave morphology restitution (TMR) index was calculated by dividing dw, here renamed as dwTW, by ΔRR (figure 2C, above):

The T-peak-to-T-end-wave morphology restitution (TpeMR) index is obtained by estimating dw restricted to the interval from the T peak to the end, dwTpe (figure 2C, below):

The Wilcoxon test (2-tailed Mann-Whitney U test) was used for continuous variables, and the chi-square test was used for categorical variables to evaluate the association of clinical (age, body mass index, RS, type 2 diabetes mellitus, hypertension, amiodarone intake, dyslipidemia, coronary artery disease, and thyroid-stimulating hormone) and ECG variables (ΔαQT, ΔαTpe, TMR and TpeMR) with the primary endpoint. Clinical variables were selected based on an a priori knowledge of their role in cardiovascular risk1 and their independence from the RS. Variables with no significant association with the primary endpoint were not included in subsequent analyses. No imputation for missing data was required, as complete information was available for all variables.

Pairwise correlations among covariates were calculated, and in high correlation (|r|> 0.8) cases, 1 variable from each correlated pair was excluded.16 Univariate Cox regression analyses were performed to determine the relationship between each clinical or ECG-based variable and the time to the primary endpoint. Hazard ratios (HRs) were calculated using the low-risk groups as the reference for categorical variables. Only the ECG-based index most significantly associated with the primary endpoint was included in the multivariate Cox regression analysis, which was performed by adjusting for the RS and all clinical variables that were significant in the univariate analysis. The proportional hazards assumption was assessed using Schoenfeld residuals, and model discrimination was evaluated using the C-index.

A value of P <.05 was considered statistically significant. Statistical analyses were performed using RStudio version 2024.12.0 (Posit PBC, United States).

Kaplan-Meier curves were derived for each variable, with a comparison of cumulative events performed using log-rank tests. For this, we dichotomized each variable by setting the optimal cutoff point as the value that jointly maximized sensitivity and specificity for risk stratification.14 From this optimal cutoff point, patients were dichotomized into 2 groups: high, if the value was above the optimal threshold, and low, if it was below the optimal threshold. Specifically for the RS, patients were stratified into risk groups based on the original publication3 as high (H), intermediate (I) and low (L) risk.

We developed a combined risk model (CRM) as the weighted sum of the RS with the variables that remained statistically significant in the multivariate Cox regression analysis, where the weights were the respective β-coefficients from the multivariate Cox model17 (see methods of the supplementary data). Then, patients were dichotomized into 2 CRM risk groups according to the optimal cutoff point and the association between the 2 risk groups and the primary endpoint was evaluated using Kaplan-Meier survival curves, the log-rank test, and univariate Cox regression analyses. Classification performance metrics, including balanced accuracy, positive and negative predictive value (PPV and NPV, respectively), sensitivity, and specificity, as well as confusion matrices, were calculated and compared between the RS and the CRM. Net reclassification improvement quantified the added predictive value of the CRM, using risk categories aligned with the primary endpoint event rate.

Sex-specific associations of the ECG-based indices in CC for risk stratification were assessed by repeating the main analysis separately in male and female patients. In addition, a sensitivity analysis was performed excluding patients receiving amiodarone treatment to assess whether its use influenced the main findings.

Finally, time-to-event analyses were explored to assess whether the predictive value of the ECG indices varied during the follow-up.

Figure 1 shows the flow diagram of the inclusion and exclusion criteria, leading to the final study population, which consisted of 144 patients with chronic CC, of whom 53 were male and 91 were female, aged 23 to 62 years. A total of 53 patients (36.8%) met the primary endpoint (21 men and 32 women) within a median follow-up period of 10.9 years. The detailed characteristics of the study population are shown in table 1.

Patients in the primary endpoint group showed higher values of RS (P <.0001) and a higher proportion of hypertension (P=.002) and amiodarone intake (P <.0001), while age, sex, body mass index, diabetes mellitus type 2, dyslipidemia, coronary artery disease and thyroid-stimulating hormone showed no significant variations. Upon comparison of ECG variables between the primary endpoint and control groups, both ΔαTpe and TpeMR were significantly higher in the primary endpoint group (P <.001 and P=.005, respectively). Clinical variables and ECG-derived indices that were not associated with the primary endpoint were not included in subsequent analyses (table 1). The AUC for RS, ΔαTpe and TpeMR was 0.913 (95% confidence interval [95%CI], 0.864-0.962), 0.693 (95%CI, 0.599-0.786), and 0.640 (95%CI, 0.541-0.738), respectively (figure S1).

Univariate Cox analysis revealed that hypertension, amiodarone intake, RS, ΔαTpe and TpeMR were all significantly associated with the primary endpoint (table 2).

Figure 3 shows the correlation coefficients among amiodarone intake, ECG-derived parameters, the RS, and hypertension. Amiodarone intake was not associated with ECG parameters, but was strongly correlated with the RS and was therefore excluded from the multivariable analysis to avoid collinearity, while the low correlations between the RS and ECG-derived parameters support their complementary and nonredundant value. Multivariate Cox analysis was then performed including ΔαTpe, as the most significant ECG-based index in the univariate analysis, hypertension and RS, all remaining significantly associated with the primary endpoint, with an HR of 13.6 (P=.020), 2.80 (P=.014), and 1.46 (P <.0001), respectively (table 2). The proportional hazards assumption was not violated for any covariates (methods of the supplementary data), and the model showed good discrimination (C-index=0.859).

Figure 3.

Matrix of correlation coefficients. The normality of continuous variables was assessed using the Shapiro-Wilk test. Spearman's rank correlation was used for nonnormally distributed continuous variables, the Phi coefficient for binary categorical variables, and the point-biserial correlation for continuous-binary associations. Coefficients were calculated based on 144 complete cases.

(0.13MB).

In line with the original publication,3 the RS was stratified as low risk, or RS(L), from 0 to 6 points, and intermediate risk, or RS(I), from 7 to 11 points. The optimal thresholds were 0.036 for ΔαTpe and 0.021 for TpeMR. Of the 144 patients studied, 95 (66%) were included in the RS(L) group and 49 (34%) in the RS(I) group. Likewise, 89 (61.8%) were included in the ΔαTpe(L) group and 55 (38.2%) in the ΔαTpe(H) group; similarly, 90 (62.5%) were included in the TpeMR(L) group and 54 (37.5%) in the TpeMR(H) group. figure 4 shows Kaplan-Meier survival curves for the low- and high-risk groups defined for ΔαTpe (figure 4A), TpeMR (figure 4B), and RS (figure 4C).

Figure 4.

Kaplan-Meier survival curves. Cumulative survival rates of individuals stratified by A: ΔαTpe ≥ 0.036 (ΔαTpe(H)), B: TpeMR ≥ 0.021 (TpeMR(H)), C: RS between 7 and 11 points (RS(I)), and D: by the 2 risk groups defined in the combined specific risk model for sudden cardial arrest. CRM, combined risk model; H, high; I, intermediate; L, low; TpeMR, T peak-to-end morphology restitution; RS, Rassi score.

(0.6MB).

The CRM was constructed by combining hypertension, RS, and ΔαTpe (methods of the supplementary data). The AUC for the CRM was 0.930 (95%CI, 0.886-0.974, figure S1). Although the CRM showed a numerically higher area under the curve (AUC) and slightly narrower CI compared with the RS, the difference was not statistically significant according to the DeLong test (P=0.15). Of the 144 patients studied, 63 (43.8%) were included in the high risk group [CRM(H)] and 81 (56.2%) in the low risk group [CRM(L)] (table S1).

Individuals classified in the CRM(H) group had a significantly higher risk of experiencing the primary endpoint (HR=23.1; P=<.0001) than those in the CRM(L) group (table 3). For reference, the HR value for participants in the RS (I) group compared with those in the RS (L) group was 17.8 (P <.0001). Kaplan-Meier survival curves for the CRM (L) and (H) risk groups are shown in figure 4D.

Figure 5A shows that the CRM correctly classified more true positives than the RS (48 vs 42) but also had more false positives (15 vs 7). figure 5B shows that the CRM had higher sensitivity (0.906 vs 0.792), and NPV (0.938 vs 0.884), while specificity (0.835 vs 0.923) and PPV (0.762 vs 0.857) were reduced, with a slight global improvement in balanced accuracy (0.870 vs 0.858). The mean net reclassification improvement for CRM vs RS was 0.542 (0.280-0.805, P <.001; see table S2).

Figure 5.

Confusion matrices (A) and classification performance metrics (B) for the current Rassi score (red) and the proposed combined risk model (CRM, blue). NPV, negative predictive value; PPV, positive predictive value.

(0.23MB).

Sex-stratified analyses showed that ΔαTpe predicted SCA-related events in both sexes, while ΔαQT showed an association only in male patients and TpeMR only in female patients. However, these associations did not persist after multivariable adjustment, and no sex-specific CRMs were supported. Similarly, in the amiodarone-free population, ECG-derived variables did not show an independent association with the primary endpoint after adjustment, whereas the RS remained the only significant predictor. Finally, a predictive performance analysis across fixed follow-up periods identified the 3-year time point as showing the best discrimination. Detailed results are provided in the methods of the supplementary data, tables S4-S8, and figure S2.

A key limitation of this study is the inability to compare results with the FIOCRUZ score20 due to the lack of multilead ECG data needed to calculate QT interval dispersion. This restricts alignment with established models and highlights the need for more comprehensive ECG data in future research. Moreover, a limitation of the CRM is that the gain in sensitivity, balanced accuracy, and NPV comes at the expense of a loss of specificity and PPV, limiting its clinical benefits. While the increased sensitivity, balanced accuracy, and NPV indicate the enhanced ability of CRM to identify patients who have been categorized as low or intermediate risk using the RS but are, nevertheless, still at high risk of SCA, this comes at the expense of reduced specificity and PPV, which can result in unnecessary monitoring or interventions in clinical practice.

Consequently, our findings need to be validated in larger populations to strengthen the reliability of our conclusions and provide insights into the adaptability of the model to diverse patient populations and clinical settings. It should be noted that this study was not blinded. In addition, interobserver reproducibility was not performed, as the measurement procedure was fully automated, ensuring observer independence and repeatability. Moreover, only a single Holter recording per patient was analyzed. When multiple recordings were available (n=52), and in view of the large mean time between recordings (3.5± 4.9 years), the recording closest to the primary endpoint was selected, as it likely contained the most relevant information about the arrhythmogenic substrate leading to the event. Consequently, intraparticipant reproducibility was considered underpowered and was therefore not assessed. Future well-powered studies with repeated ECG recordings are needed to evaluate the intraparticipant reproducibility of our results.