Discussion

Although it has been extensively investigated, the relationship between obesity and clinical outcomes for patients with established CAD remains controversial. This meta-analysis which includes 89 studies and more than 1.3 million patients, represents the largest analysis on this subject. Our results demonstrated a J-shaped relationship between BMI and all-cause mortality among patients with CAD, which was attenuated with longer follow-up. Using normal-weight as the reference, being underweight is consistently associated with a higher risk of mortality, while being overweight is consistently associated with a lower risk of mortality. The presence of obesity was associated with a lower risk of mortality but the benefit disappears after 5 years of follow-up. Additionally, grade II/III obesity tends to be associated with a lower risk of short-term mortality but a higher risk of mortality at more than 5 years of follow-up. The similar J-shaped pattern was also seen for cardiovascular mortality, and across different types of treatment strategies.

The inverse relationship between BMI and mortality in our findings contrasts with prior individual studies that have reported no relationship between BMI and mortality,9–11 but are largely consistent with a prior meta-analysis by Romero-Corral et al.17 By summarising the data from 40 studies including 250 000 patients with at least 6-months of follow-up, Romero-Corral et al found that the adjusted RRs of all-cause mortality were 0.87 (95% CI 0.81 to 0.94) for overweight and 0.93 (95% CI 0.85 to 1.03) for obesity relative to normal-weight. While the BMI cut-off points across individual studies enrolled in Romero-Corral et al's meta-analysis demonstrated wide variability, we only included studies using standard BMI classifications. Our study also provides a much larger sample size and a broader range of follow-up, ranging from inhospital to over 12 years of follow-up. Nevertheless, our analysis demonstrated relatively similar risks of long-term mortality in overweight patients as compared with Romero-Corral et al (RR 0.82 95% CI 0.74 to 0.90), and while we examined obesity in a graded fashion in which patients with less severe obesity (grade I) and more severe obesity (grade II and III) were evaluated separately, the effect we demonstrated of grade II/III obesity on mortality was quite similar to the prior analysis as the RR was 0.93 (95% CIs 0.82 to 1.05). Therefore, while our analyses take a more nuanced approach at the effect of obesity on long-term mortality in patients with CAD, our findings are consistent with prior analyses. Although a high level of variability was found across individual studies, subgroup analyses and meta-regression suggested that the J-shaped relationship pattern we demonstrated is consistent across all study and patient subsets.

Despite the overwhelming evidences and consistent results across subgroups, we should be cautious to draw the conclusion that excess body fat is a protective factor in patients with established CAD. Like many of these individual studies, our analysis is susceptible to multiple biases and to confounding.

Possible explanations for the ‘obesity paradox’ have included earlier presentation of obese patients,18 questionable validity of BMI to adequately account for body fat19 ,20 and the inability of BMI to account for a patient's overall level of fitness.21–23 A recent study24 found that smoking status is an important effect modifier of BMI on mortality in patients with diabetes, demonstrating a linear relationship among non-smokers but a J-shaped relationship among smokers. In order to examine the influence of confounding and provide precise estimates, we performed separate analyses for adjusted estimates. The results presented herein provide little support for the suggestion that the baseline confounding factors are important causes of bias. Although the effect sizes of mortality appeared to be smaller in adjusted studies, the favourable effect of being overweight or obese versus normal-weight were consistent across unadjusted and adjusted estimates. While separate analyses did not find substantial differences between smoking-adjusted and overall estimates, indicating no significant smoking-related bias in this study, no individual study to date has performed subgroup analysis among patients who have never smoked; it is thus not possible for our study to examine the residual bias related to smoking. Additionally, treatment strategies, especially revascularisation, have been reported to be a potential source of bias.17 However, we did not find significant effect modification by treatment strategies in meta-regression and subgroup analyses performed, and sensitivity analyses did not find significant modification of effect size or reduction of variability after excluding any kind of treatment strategies.

Another important concern is reverse causation. In analyses using baseline weight as the exposure, low body weight could be a result of weight loss due to more severe illness or underlying chronic disease. In a post hoc analysis of three trials,25 significant weight loss (loss >0.1 kg/baseline BMI) within the first 3 months to 1 year after acute MI was found to be an independent predictor of mortality. Exclusion of subjects with known illness at baseline and censoring of short-term data are routine methods to reduce this bias.24 ,26 Our meta-analysis is susceptible to reverse causation because very few studies took into account body changes prior to the baseline weight at study enrolment, and almost all the studies involved patients who had a prior history of CAD, which substantially increases the susceptibility to reverse-causation bias. Additionally, the attenuated benefit of overweight and obese patients in long-term versus short-term follow-up demonstrated in this analysis may also be an indicator of reverse causation, with the assumption that short-term follow-up data is more likely to be influenced by pre-existing illness that could lead to a lower baseline body weight.

The main strength of this meta-analysis is the large sample size and number of studies included, which make our findings robust as compared with prior studies on this issue. Limitations of our meta-analysis include the fact that individual patient data were not available for these meta-analyses, and therefore the variables in the adjusted risk estimates are different across studies, which thus limit the capacity to fully explore effects in subgroups. Most contributing studies only provided baseline BMI values, so it is not possible for this meta-analysis to take into account the potential influence of weight change over time on prognosis. Risk estimates also largely varied across individual studies (risk ratios, HRs, and ORs, and raw event number). HRs can take into account time of events and hence are considered the best estimate for follow-up data. In the present analysis, we also performed separate analyses confined to HR, and found that the results are consistent with the main results (see online supplementary eFigure 10), indicating the robustness of our findings. Second, although we used a strict BMI classification as the inclusion criteria, heterogeneity was present almost among all analyses. Attempts to investigate sources of heterogeneity through various sensitivity analyses and meta-regression did not find a simple explanation or method to account for this variability. The inconsistency of follow-up duration across individual studies may partly explain the heterogeneity of long-term analyses, but a high level heterogeneity still exists within different follow-up subgroup analyses. Substantial heterogeneity has also been found in a prior meta-analysis on this issue,17 which was attributed to sufficient statistical power to detect small differences across studies due to large numbers of events and a large sample size by the authors. Finally, only BMI was evaluated in this analysis. Other modalities associated with body weight, including body composition27 ,28 and fitness,21–23 ,27 were not included due to insufficient data on these parameters.

Rather than providing new evidence regarding the ‘obesity paradox’, this study reminds us of the limitations of current literature and reinforces the importance of better designs for future studies. To interpret the true causal relationship between body weight and prognosis in patients with established CAD, multiple biases and confounders, particularly reverse causation, smoking status and fitness, should be taken into account carefully. Long-term follow-up is strongly recommended due to the susceptibility to reverse-causation bias of shorter follow-up. Prospective randomised controlled studies are needed to assess the survival benefit of purposeful weight-control programmes across different BMI categories in patients with CAD. Until more definitive persuasive evidence is available, the current results should not deter the enthusiasm of weight control through exercise and diet for patients with CAD, especially for those with excess body weight.