Distinct implications of body mass index in different subgroups of nonobese patients with heart failure with preserved ejection fraction: a latent class analysis of data from the TOPCAT trial

This is a post hoc analysis of the TOPCAT trial. We obtained the data from the Biologic Specimen and Data Repository Information Coordinating Center of the US National Heart, Lung, and Blood Institute via an approved proposal. TOPCAT was a multicenter, international, randomized, double-blinded placebo-controlled trial, which enrolled 3,445 adults with HFpEF from over 200 clinical centers. The aim of the trial was to investigate whether spironolactone could improve prognosis in patients with HFpEF. The design of TOPCAT has been published in detail previously [16]. Briefly, eligible patients were required to have: an age of ≥ 50 years; at least one symptom and one sign of heart failure; left ventricular ejection fraction of ≥ 45%; controlled systolic blood pressure; serum potassium of < 5.0 mmol/L; elevated natriuretic peptide levels within the last 60 days or at least one HF hospitalization in the last 12 months. The trial was approved by the ethics committee at each study center. Every participant signed an informed consent form.

In our study, we excluded patients enrolled from Russia and Georgia (n = 1678), for the regional difference in TOPCAT [17]. Patients with missing data on height or weight (n = 9) were also excluded. Finally, a total of 1758 patients were analyzed in our study.

Obese and nonobese patients were defined by a BMI cutoff value of 30 kg/m², and LCA was performed among the nonobese patients (n = 623). For the LCA variables, two steps of selection were performed. The first step was manual selection. The inclusion criteria in this step were not stringent to avoid subjective bias and missing unrecognized clustering variables. Variables about demographic features, medical history, physical examination, and laboratory investigation potentially related to HFpEF prognosis or pathophysiology were eligible. Two junior cardiologists (BD and XH) screened the variable dictionary of the TOPCAT dataset, and potential candidates were adjudicated by a senior cardiologist (CL). Thirty-six variables were manually selected and are listed in Additional file 1: Table S1. The second selection step was algorithm-based. A backward stepwise algorithm was adopted to discard redundant and noninformative variables because including these variables in the LCA model will negatively affect the clustering performance [18]. In brief, all of the 36 variables were included in the LCA model. The algorithm assessed the usefulness of each variable and discarded the least useful one at each step. To avoid missing potential clustering variables, the algorithm reevaluated the usefulness of all discarded variables after each removal and tried to add the most likely useful one back to the model. The action of “removal” or “adding” would be accepted or rejected depending on whether it improved the performance of the model assessed by the difference in the Bayesian Information Criterion (BIC). This calculation iterated until no further action was accepted. The abovementioned selection was achieved using the LCAvarsel package in R. The process of algorithm-based variable selection is presented in Additional file 1: Table S2. Finally, 10 variables namely race, Kansas City Cardiomyopathy Questionnaire (KCCQ) overall summary score, previous HF hospitalization, diabetes, diastolic blood pressure, diuretic usage, beta-blocker usage, hemoglobin, albumin and glomerular filtration rate (GFR) were selected for the LCA model (Additional file 1: Table S3). The optimal number of subgroups was determined by comprehensively evaluating BIC, Akaike information criterion (AIC), G², χ², adjusted BIC (aBIC), and consistent AIC (cAIC). Three out of the six indexes supported the optimal group number to be two (Additional file 1: Table S4). Partial probabilities of the phenotype membership for the selected variables are summarized in Additional file 1: Table S5. The probability of a patient belonging to a certain group was determined by multiplying partial probability of each variable. Finally, the patient was assigned to the group that had the highest probability.

The outcomes in our study included all-cause mortality, cardiovascular mortality, and a composite outcome of cardiovascular mortality, aborted cardiac arrest, or HF hospitalization defined as primary by TOPCAT. Noncardiovascular causes accounted for a larger proportion of mortality in HFpEF compared with that in heart failure with reduced ejection fraction [19, 20]. Therefore, noncardiovascular mortality was also evaluated. All the outcomes were adjudicated by a clinical end-point committee at Brigham and Women’s Hospital.

Comparisons were made between the two nonobese groups and the obese group. The continuous variables are presented as mean ± standard deviation or median with the interquartile range depending on their normality. The categorical variables were described using frequencies with percentages. Differences in baseline characteristic data among the three groups were compared using a one-way ANOVA or Kruskal-Wallis test for continuous variables. Bonferroni’s test was applied for pairwise comparisons in the post hoc analysis if the results were significant. The categorical variables were compared using chi-squared test. The Kaplan-Meier survival curve with a log-rank test was used to show the differences in survival among the groups. Cox proportional-hazards regression model was performed to evaluate the association of different phenotypes with all-cause mortality and the composite outcome, and the obesity group served as the reference group. For cardiovascular and noncardiovascular mortality, Fine and Grey’s competing risk regression model was used to adjust for the competing risk from each other.

To evaluate whether the prognostic implications of the nonobese subgroups were independent of demographic differences and randomized treatment, we used two methods to adjust for the potential confounding effect of age, gender, country of origin, ethnicity, and randomized treatment. The first was the multivariate Cox or competing risk regression model including both grouping and the above variables. The second was assigning weights to each individual according to their propensity score to be obese using standardized mortality ratio weighting [21, 22]. The rationale of the weighting process was to ensure that the weighted estimates of the abovementioned characteristic were comparable among the three groups, and therefore, the corresponding weighted regression model should be free from the confounding effects of these variables.

Spironolactone treatment was shown to have a differential effect on the different subgroups of patients with HFpEF [23]. A sensitivity analysis restricted to the placebo arm was performed using the multivariate regression model.

The baseline characteristics of the groups were summarized in Table 1.

The patients with physiological non-obesity had a median age of 76, a male proportion of 57.19%, and a median BMI of 26.85 kg/m². Only 23.86% of the patients were classified as NYHA III/IV. The median KCCQ overall score for quality of life was 75. For metabolic comorbidities, 85.21% had hypertension, 65.14% had dyslipidemia, and 13.73% had diabetes. These patients were relatively active with a median of 4 metabolic equivalents of task (METs) per week. Approximately 76.84% of patients relied on diuretics to relieve HF syndromes. Median GFR was 68.46 ml/min/1.73m². For the nutritional indexes, the mean hemoglobin concentration was 14.09 g/dL and the median serum albumin concentration was 4.1mg/L.

The median age of the patients with pathological non-obesity was 78, and 47.34% were male. The median BMI was 26.08, which was comparable to the Physiological non-obesity group. Approximately 31.07% of them were classified as NYHA III/IV. The KCCQ overall score was lower than patients with physiological non-obesity, with a median of 62.24. The prevalence of hypertension (86.39%) and dyslipidemia (66.57%) were comparable to the Physiological non-obesity group, but the prevalence of diabetes was much higher (41.13%). The patients were less active with a median of 2.3 METs per week. A higher proportion of them relied on diuretics for HF (89.05%). Renal function was worse with a median GFR of 55.28 mL/min/1.73 m². The hemoglobin (average 11.86 g/dL) and serum albumin (median 3.8 mg/L) concentrations were also lower than patients with physiological non-obesity.

After 2 years of follow-up, patients with pathological non-obesity were more likely to experience weight loss of ≥10% body weight compared to patients with physiological non-obesity (11.34% vs. 4.19%, P = 0.009) (Fig. 3).

Patients with obesity were relatively younger with a median age of 69, and 49.07% of them were males. The median BMI (36.57 kg/m²) was high. Approximately 39.31% of this group was classified as NYHA III/IV, which was higher than that of both nonobese groups. The quality of life was worse than that in the nonobese groups with a median KCCQ overall score of 54.69. Metabolic syndromes were the most prevalent, with 92.25% of hypertension, 73.74% of dyslipidemia, and 53.57% of diabetes. The patients were also less active compared to patients with physiological non-obesity with a median of 1.5 METs per week; 92.15% of them relied on diuretics for treatment. GFR, hemoglobin concentration, and serum albumin concentration lay between the two nonobese groups (median GFR = 60.86 ml/min/1.73m², mean hemoglobin = 12.84 g/dL, median albumin = 3.9 mg/L).

For the overall cohort, 383 (21.8%) patients died, 222 of them died of cardiovascular causes, and 161 of them died of noncardiovascular causes. Five hundred and nineteen (29.5%) patients experienced a composite outcome. The crude rates of these outcomes were compared among the 3 groups. Patients with pathological non-obesity had the highest all-cause, cardiovascular, and noncardiovascular mortality rates among the three groups. Patients with physiological non-obesity and obesity had comparable mortality rates, but the patients with physiological non-obesity had a lower rate of composite outcome (Table 2). The Kaplan-Meier survival curve and the unadjusted regression model showed similar trends (Fig. 4).

To adjust for the potential confounding effects of the demographic differences in age, gender, country of origin, ethnicity, and randomized treatment, the multivariate regression, and regression with standardized mortality ratio weighting were performed.

In the multivariate regression model, compared to patients with obesity, patients with physiological non-obesity had a 47% decrease in risk of HF composite outcome (HR 0.53, 95% CI 0.40–0.70, P < 0.001) and a trend of lower all-cause mortality risk (HR 0.75, 95% CI 0.55–1.01, P = 0.06). In contrast, patients with pathological non-obesity were associated with a 59% increase in all-cause mortality risk (HR 1.59, 95% CI 1.24–2.02, P < 0.001), which appeared to be driven by an increased cardiovascular mortality risk (HR 1.64, 95% CI 1.19–2.25, P = 0.003) (Fig. 4).

Using standardized mortality ratio weighted models yielded similar results. Age, gender, country of origin, ethnicity, and randomized treatment were well balanced after weighting (Additional file 1: Table S6). Patients with physiological non-obesity were associated with a lower risk of all-cause mortality (HR 0.70, 95% CI 0.49–1.00, P=0.048) and HF composite outcome (HR 0.50, 95% CI 0.36–0.70, P<0.001), while patients with pathological non-obesity were associated with a higher risk of all-cause (HR 1.50, 95% CI 1.11–2.03, P=0.009) and cardiovascular (HR 1.52, 95% CI 1.03–2.24, P=0.03) mortality (Fig. 5).

The sensitivity analysis restricted the multivariate models in the placebo arm and yielded similar results to the analysis in the overall population (Table 3), except that some differences were no longer significant due to a reduced sample size.

Several limitations should be taken into consideration. First, our analysis was restricted to HFpEF. It is not clear whether these findings could be generalized to HFrEF. Second, this was a hypothesis-generating study without external validation. Our findings need to be validated by future studies. Third, the sample size, especially for the nonobese group, was limited. Fourth, the history of obesity and body weight change before trial recruitment could provide more useful insights, but this information was not available. Fifth, the cachexic nature of patients with pathological non-obesity was inferred from the clinical manifestation and prognosis. A more comprehensive nutritional evaluation or assessment of the adipose tissue and skeletal muscle mass was not available to confirm the existence of cachexia. Sixth, the difference biological effect of visceral adipose tissue and subcutaneous adipose tissue might contribute to the obesity paradox. However, this information was not available in the TOPCAT dataset. The fat distribution in different subgroups of nonobese HFpEF needed to be evaluated in future studies. Seventh, the present analysis mainly focused on nonobese patients to explain the “obesity paradox.” However, obese patients with HFpEF could also be heterogeneous as BMI did not reflect adipose distribution in the body.

There were also strengths in this study. First, the LCA model comprehensively assessed multiple features of the nonobese HFpEF population, which was better in terms of subgrouping compared with the traditional subgroup analysis based on a single variable. In addition, the use of an algorithm-based LCA variable selection further reduced the subjective bias in subgrouping the nonobese patients with HFpEF. Furthermore, the prognostic difference was confirmed using separate analysis with different methods to adjust for potential demographic confounders and a sensitivity analysis with only the placebo arm.