The Women’s Health Initiative (WHI) study recruited 161,808 post-menopausal women with ages 50–79 years from 1993 to 1998 at 40 clinical centers in United States. Details of recruitment, baseline assessments and follow-up have been published previously [20–22]. Briefly, the WHI included a cohort of 93,676 women in a prospective observational study (OS) and 68,133 women in one or more of the three clinical trials (CT): hormone therapy (HT); calcium and vitamin D; or dietary modification trial. In 2010, a sub-cohort (HF cohort) of 44,174 women including all women who participated in the HT, and oversampled for African American and Hispanic/ Latina women from both the CT and OS arms of WHI, were evaluated retrospectively and prospectively until Feb 28, 2020, for incident hospitalized HFpEF and HFrEF events by trained physician adjudicators [23]. The primary analysis included participants from this HF cohort of women in whom data collection on cardiac imaging and other tests to define HF subtypes was performed after the conclusion of the main WHI study. We have compared this sub-cohort of WHI participants to those in the overall cohort for CHD and HF surveillance and found no substantial differences in CHD or HF risk factors [24]. Participants without baseline ECG on their initial assessment, or participants with ECG showing artifacts, ectopic beats, arrhythmias or second/ third-degree conduction blocks and noise (defined as < 80% normal RR intervals) were excluded from analysis. As coronary ischemia and anti-arrhythmic medications may affect HRV, we excluded those with self-reported history of CHD (prior history of MI, percutaneous transluminal coronary angioplasty, or coronary artery bypass grafting) and those on anti-arrhythmic medications from the sample. The final sample for analysis included 28,603 participants after excluding those with self-reported history of HF at baseline to have a disease-free incident HF cohort (Fig 1). Informed written consent was obtained from study participants at each participating center. All WHI studies were approved by the research ethics committee at each participating center. Given that this project used only deidentified data from WHI, it met criteria for exemption by the Providence Veterans Affairs Medical Center Institutional Review Board.

Centrally trained and officially certified technicians performed digitally recorded 10-second standard 12-lead electrocardiogram (ECG) in a resting, supine or semi-recumbent position at enrollment in the WHI. Technicians used comparable procedures to record ECGs with MAC PC electrocardiographs (GE Marquette Electronics, Inc., Milwaukee, WI). Recorded ECGs were telephonically transmitted to a central laboratory (Epidemiological Cardiology Research Center, Wake Forest School of Medicine, Winston Salem, NC) for identification of technical errors, inadequate quality, and analysis using the 2001 version of the Marquette 12-SL program (GE Marquette, Milwaukee, WI). Electronic reading of a 10-second ECG produced two measures of the cardiac ANS activity: 1) the root mean square of successive differences in normal-to-normal RR intervals (RMSSD, ms) = [[Σ(RRj +1 –RRj)2]/n]0.5, and 2) the standard deviation of all normal-to-normal RR intervals (SDNN, ms) = [[Σ(RRmean–RRj)2]/(n − 1)]0.5 [25]. The median resting heart rate was measured in beats/ min on a 12-lead ECG. The repeatability and accuracy of short-term, time domain measures of HRV, i.e, SDNN and RMSSD have been described previously [26].

Primary outcome of interest was incident hospitalized HF and its subtypes; HFpEF and HFrEF. Incident hospitalized heart failure was ascertained annually by medical record abstraction of all self-reported hospitalization, by trained adjudicators using the standardized methodology as previously described (See Appendix in S1 File) [23]. This process allowed adjudicators to define acute decompensated HF, chronic HF, and unclassifiable or unknown event (no HF). Acute decompensated HF was further classified into HFpEF, HFrEF and unknown ejection fraction HF (HFuEF). HFrEF was defined as HF with an EF < 50% and HFpEF was defined as HF with an EF ≥ 50%. The acute HF classification system used in this analysis has been shown to have good agreement with other epidemiological HF algorithms [27].

The following potential confounders and covariates determined based on prior knowledge were included for adjustment in the models; age, race, ethnicity, income, education, hypertension, diabetes mellitus, smoking status, alcohol intake, hyperlipidemia, body mass index (BMI), physical activity, left ventricular hypertrophy (LVH), use of beta-receptor blocking agents, calcium-channel blockers and hormone therapy use. The methods of baseline data collection in WHI [20–22] have been published elsewhere. Study questionnaires, physical measurements, and quality assurance have been detailed previously [20, 21]. Race and ethnicity were self-reported in the baseline questionnaire. Participants underwent measurement of blood pressure, height, weight, and hip and waist circumferences at enrollment. BMI was calculated as weight (in kilograms) divided by the square of measured height (in meters squared). Age, income, education, history of hypertension, diabetes, hyperlipidemia, smoking, and alcohol use were ascertained by self-report on baseline questionnaires. Participants with measured resting systolic blood pressure ≥140 mm Hg or diastolic blood pressure ≥90 mm Hg at the initial clinic visit were also classified as hypertensive. Physical activity was assessed using self- reported questionnaires and frequency, intensity, duration, and types of physical activity were evaluated and converted to MET-hrs/week as previously described [28, 29]. LVH was determined by voltage criteria using Minnesota ECG Code [30].

Baseline characteristics of study participants were categorized according to lowest (Q1) and highest (Q4) quartiles of RMSSD, SDNN and rHR. Continuous variables were presented as means ± SD and categorical variables as proportions. Differences in baseline characteristics between Q1 and Q4 of RMSSD, SDNN and rHR were tested by chi-square for categorical variables and ANOVA for continuous variables.

We used Cox proportional hazards regression analysis to estimate hazard ratios and 95% confidence intervals (CI) of incident hospitalized HF, HFrEF, and HFpEF for each quartile of RMSSD, SDNN and rHR, using the highest quartile (Q4) of RMSSD and SDNN, and the lowest quartile (Q1) of rHR as a reference. Separate Cox-proportional hazards analyses were performed to estimate hazard ratio for incident HF, incident HFpEF and incident HFrEF. Follow-up time for each participant was calculated from the date of study enrollment to the date of a confirmed incident HF event, last follow-up known to be without HF, death from any cause, or until Feb 2020 when follow-up ended, whichever came first. The proportional hazards assumption was checked by Schoenfeld residuals performed on all the variables used in the model. Potential confounders determined based on prior knowledge were included in the multivariable Cox-regression model. Age, hypertension, diabetes mellitus, LVH, BMI, and total recreational physical activity were considered as potential confounders. As such, we used 3 sequential models with model 1 adjusting for age and race/ ethnicity, and model 2 additionally adjusted for smoking status, alcohol, physical activity, BMI, education, hypertension, diabetes, hyperlipidemia, LVH, use of beta-receptor blocking agents, calcium-channel blockers and hormone therapy use. Model 3 was used to adjust for time-varying incident CHD. Incident CHD was defined as the first occurrence of fatal or non-fatal myocardial infarction, percutaneous transluminal coronary angioplasty, or coronary artery bypass grafting. CHD events were self-reported annually and then centrally adjudicated by trained physicians after obtaining medical records and/or death certificates. In addition, we performed sensitivity analysis by excluding first two years of follow-up and participants on loop diuretics at baseline to further minimize reverse causation and the possibility of undiagnosed HF at enrollment, respectively. We also adjusted for mortality as a competing risk of death in supplementary analysis. Trend testing across quartiles of RMSSD, SDNN and rHR was calculated using the median value within each quartile. Alpha error for two-sided hypothesis tests was set at 0.05. Cumulative HF free survival curves were plotted using Kaplan-Meier survival analysis and log-rank test was used to compare differences between quartiles of SDNN and RMSSD. All analyses were conducted using SAS version 9.4 (SAS Institute Inc. Cary, NC).

At baseline, the overall mean age of participants was 62.6±7.1 years with race and ethnicity being predominantly non-Hispanic White (69%) followed by African-American (21%) and Hispanic women (9%) (p<0.001). Compared to lowest quartile (Q1), study participants were younger in the highest quartile (Q4) of SDNN and RMSSD. Among the quartiles of SDNN, RMSSD and rHR, the median SDNN in Q1 was 7.5 ms (IQR: 0.91–10.19 ms) compared to 33.2 ms (IQR 24.1–313.2 ms) in Q4, median RMSSD in Q1 was 7.7 ms (IQR:1.2–10.5 ms) compared to 36.4 ms (IQR: 25.6–361.23 ms) in Q4 and median rHR in Q1 was 55 bpm (IQR: 33–59) compared to 78 bpm (IQR: 73–123) in Q4. The prevalence of diabetes mellitus, hypertension, hyperlipidemia, and LVH was higher among women in Q1 of SDNN and RMSSD compared to Q4 and in Q4 of rHR compared to Q1. Women in Q1 of SDNN and RMSSD, and Q4 of rHR were physically less active, had higher alcohol consumption and had a lower proportion with college education than women in Q4 of SDNN and RMSSD, and Q1 of rHR. (Table 1).

Pearson correlation coefficients between SDNN, RMSSD and rHR showed a strong positive correlation between RMSSD and SDNN (r = 0.90, p<0.001) whereas a weaker inverse correlation was found between rHR and SDNN (r = -0.28, p<0.001), and rHR and RMSSD (r = -0.32, p<0.001).

During a median follow-up of 17.6 years, 2022 overall HF events occurred, with incidence rate noted to be greatest with lowest quartiles (Q1) of RMSSD and SDNN (Figs 2 and 3), and highest quartile (Q4) of rHR (Table 2). In a cox-proportional hazards regression analysis, low HRV (RMSSD and SDNN) and as well as high rHR were associated with significantly higher risk of total overall HF (Q1 SDNN compared to Q4 SDNN: 1.22, 95% CI 1.07, 1.39; Q1 RMSSD compared to Q4 RMSSD: 1.17, 95% CI 1.02, 1.33; Q4 rHR compared to Q1 rHR: 1.31, 95% CI 1.15, 1.50) (Table 2). A total of 1025 incident HFpEF and 654 incident HFrEF events occurred during a median follow up of 17.6 years. On analyzing the association of HRV with HF subtypes, low HRV (SDNN) was associated with elevated risk of HFpEF, with the greatest risk found in the Q1 which remained significant in a fully adjusted model 3 (Q1 SDNN compared to Q4 SDNN hazard ratio: 1.22, 95% CI 1.02,1.47), whereas no association of HRV was found with HFrEF risk (Tables 3 and 4). High rHR was associated with elevated risk of both HFpEF and HFrEF, with the greatest risk found in the Q4 (Q4 rHR compared to Q1 rHR, HFpEF: hazard ratio = 1.34, 95% CI 1.1,1.62 and HFrEF: 1.33, 95% CI 1.05,1.68) (Tables 3 and 4).

In a sensitivity analysis where we excluded the first two years of follow-up and participants on loop diuretics at baseline, and additionally adjusted for mortality as a competing risk, the association of low SDNN remained statistically significant with elevated risk of HF (hazard ratio for Q1 SDNN: 1.17, 95% CI 1.02, 1.35) but not with HFpEF (hazard ratio for Q1 SDNN compared to Q4: 1.18, 95% CI 0.97, 1.43), whereas high rHR remained statistically significantly associated with elevated risk of overall HF (hazard ratio for Q4 rHR compared to Q1: 1.22, 95% CI 1.06, 1.40) and HFpEF (hazard ratio for Q4 rHR: 1.26, 95% CI 1.04, 1.54) but not with HFrEF (hazard ratio for Q4 rHR: 1.23, 95% CI 0.96, 1.57). The association of low RMSSD became statistically insignificant with elevated risk of HF (Q1 RMSSD compared to Q4: hazard ratio 1.12, 95% CI 0.98, 1.27) (See S1 File supporting information–contains all the supporting tables)