Using individual-level data from the UKBB and HUNT2 cohorts, we performed one-sample and factorial MR analyses to investigate the causal effects of individual sleep traits (insomnia symptoms, sleep duration and morning chronotype) and their joint effects on the risk of AMI. We found evidence of an adverse causal effect of insomnia symptoms and a weak causal effect of short sleep on the risk of incident AMI, while long sleep had a protective effect in UKBB. We found no statistical evidence of interaction effects between sleep traits on the risk of AMI, but those with a high genetic risk for two sleep traits in combination (including insomnia symptoms, short sleep, and a morning chronotype) had the highest risk of AMI in UKBB. Moreover, our results showed a protective effect of genetically predisposed long sleep that was not affected by additionally being genetically predisposed to insomnia symptoms or a morning chronotype on incident AMI in UKBB. However, these results were not replicated in HUNT2, where the estimates were imprecise. These findings indicate that the main effects of sleep traits on the risk of AMI are likely to be independent of each other.
Potential mechanisms
The underlying mechanisms by which insomnia symptoms or short sleep increase in the risk of AMI are multifactorial [
65]. Insomnia and short sleep independently increase the risk of autonomic dysfunction, by increasing sympathetic tone (stress response) consequently accompanied by increased metabolic rate, increased heart rate, and decreased heart rate variability [
66‐
69]. Furthermore, experimentally induced sleep restriction has been shown to cause hormonal imbalance which stimulate proinflammatory pathways [
70], increase appetite [
71,
72], and increase insulin resistance [
73]. These autonomic and hormonal disturbances lead to hypertension [
74,
75], diabetes [
73], dyslipidemia, and obesity [
71,
72], thus constituting a set of interrelated metabolic disorders that are pathophysiological in the development of cardiac dysfunction by accelerating endothelial dysfunction and atherosclerosis [
76].
Our findings and these potential mechanisms might raise a concern that insomnia symptoms and short sleep could be regarded as similar traits. However, insomnia symptoms and sleep duration were found only moderately phenotypically (
r = −0.25;
P < 0.001) and genetically (
rg = −0.50;
P < 6×10
−17) correlated to each other [
77]. It is also important to highlight that our findings on the joint causal effects of insomnia symptoms and short sleep on the risk of AMI do not employ that concomitant presence of insomnia symptoms and short sleep causes higher increase in risk of AMI through overstimulation of the suggested underlying mechanisms, or involve any supplementary mechanisms yet to be determined.
The underlying mechanism by which chronotype may influence AMI is not yet established. Studies have found evening chronotypes have more susceptibility for cardiometabolic risk behaviors and risk factors [
12,
78,
79]. On the contrary, our causal findings suggesting that having a morning chronotype may be detrimental for incident AMI compared to having an evening chronotype might be explained by the concomitant genetic predisposition to insomnia symptoms or short sleep.
Strengths and limitations
This MR study leverages genetic information to assess the causal relationships between sleep traits and AMI, reducing the potential bias due to residual confounding, reverse causation, and measurement error in conventional observational studies [
22]. The novelty of this study is our application of factorial MR to explore the causal interplay between sleep traits on the risk of AMI, where participants were grouped based on their genetic predisposition for multiple sleep traits [
60]. We are not aware of another study that has investigated the joint effects of sleep traits in the MR context. Another major novelty is that the study benefitted from the use of results from three large GWASs for insomnia symptoms [
30], sleep duration [
31], and chronotype [
32] and used two large cohorts (UKBB and HUNT2) to replicate the findings. Moreover, this study draws on the principle of triangulation [
80], where findings were compared from different methodological approaches, which further strengthened evidence supporting causation.
Nonetheless, there are a number of limitations of this study. Factorial MR analysis is usually underpowered to detect interaction which may raise the concerns of false negative results [
60]. However, this study included the UKBB cohort with 332 676 participants constituting the largest factorial MR study on sleep traits to date. The strong instrument strength observed in UKBB cohort partially overcomes concerns due to underpowered factorial MR findings [
81]. Another limitation is that although factorial MR can identify whether two independent exposures interact and have a joint effect of public health importance [
81], it assumes exposures remain stable throughout the life course. Thus, the magnitude of effects should be cautiously interpreted.
Also, the validity of MR findings can be weakened by pleiotropy [
82]. We used several sensitivity analyses to investigate possible sources of bias in MR. We found that the genetic risk for insomnia symptoms was strongly associated with BMI, smoking status, depression, and education among other covariates [
30], which may be indicative of confounding, mediation, or horizontal pleiotropy. Further to this, our results remained consistent across various MR methods, except for insomnia symptoms which showed evidence of an unbalanced pleiotropy in MR-Egger analysis. Additionally, previous studies have shown only mild attenuation of causal effects of insomnia symptoms on CAD risk when adjusted for BMI, smoking, depression, and education using multivariable Mendelian randomization (MVMR) [
25,
26]. Moreover, simulations have shown that MR-Egger may be unreliable when applied to a single dataset [
59], and this is a limitation of our study.
The sleep traits were based on self-report. It remains unclear if self-reported sleep duration represents time in bed or actual sleep time. Also, the insomnia questions in UKBB or HUNT2 did not cover all aspects of insomnia (difficulty falling asleep, night awakenings, waking up early and daytime impairments) [
83]. Chronotype in this study was assessed from a single question in UKBB, whereas validated instruments such as the Morningness-Eveningness Questionnaire and the Munich Chronotype Questionnaire use diverse questions to better estimate chronotype [
84,
85]. Other sleep traits (e.g., sleep apnea, snoring, daytime napping) were not included, and we do not know whether these interact with insomnia symptoms or sleep duration. Moreover, the sleep traits we used are binary exposures (except for 24-hour sleep duration), which are likely coarsened approximations of the true latent exposure [
86]. This opens up alternate pathways from the genetic instruments to the outcome, which may violate the exclusion restriction assumption, resulting in biased effect estimates [
86]. In addition, causal estimates from MR of binary exposures on a binary outcome are difficult to interpret [
87].
Due to the small sample size in HUNT2, we might have missed weak causal effects due to insufficient power. In addition, the genetic instrument explained little variance in short sleep and long sleep within HUNT2, implying possible weak instrument bias [
88] and leading to wide CIs as shown in the bootstrap simulations [
89]. Furthermore, SNPs for short and long sleep were not replicated in other independent cohorts [
31], meaning that the GRS used is not validated in any other population.
The inclusion of UKBB in all exposure GWASs could lead to winner’s curse that might bias the causal estimates in UKBB [
90]. We therefore used unweighted GRS for our exposures in UKBB as recommended [
51]. Also, we derived GRS for insomnia symptoms and chronotype composed of SNPs that replicated in an independent study (23andMe) [
30,
32], which showed similar estimates, indicating winner’s curse is unlikely to have substantially biased effect estimates. However, we could not apply the same approach to explore the impact of winner’s curse on the sleep duration due to the limited sample size of the replication datasets in those studies [
31], meaning that genetic associations might be imprecise.
The variation in the occurrence of AMI between UKBB (2.35%) and HUNT2 (10.03%) may be attributed to several factors related to the composition of the cohorts: (a) the HUNT2 cohort followed up relative older participants, aged 20 years or above, with a mean baseline age of 48 years, while UKBB consisted of participants aged 40 to 69 years, with a mean baseline age of 56 years; (b) the duration of follow-up was longer in HUNT2, spanning 20.4 years, compared to UKBB’s follow-up period of 11.7 years; (c) UKBB (5.5% response rate) may represent a healthier sample [
91], whereas HUNT2 (69.5% response rate) may be a more representative sample [
36]; and (d) baseline differences in the two underlying populations or differences due to time trend (for example, more current smokers in HUNT2 which was conducted about a decade earlier than UKBB). Moreover, competing risk from death among participants would potentially hinder the occurrence of AMI, that might overestimate the risks [
92]. This is another limitation of our study.
Finally, our findings rely on analyses in UKBB due to its large sample. However, the generalizability of these findings may be limited due to a selected sample (5.5% response rate) in the UKBB cohort, which can bias both observational and MR estimates [
93,
94]. Selection bias may artificially induce associations between genetic variants and confounders leading to the instrumental variable becoming invalid [
95]. This might partly explain differences in UKBB and HUNT2 estimates observed in this study, where HUNT2 sample (69.5% response rate) more closely represents target population. The difference in demographics of the two cohorts might also cause inconsistent estimates. Moreover, the inclusion of cohorts from the European ancestry may further restrict generalizability of our findings.