Sudden cardiac and sudden unexpected death related to antipsychotics: A meta‐analysis of observational studies

To estimate the risk of sudden cardiac death (SCD) or sudden unexpected death (SUD) related to individual antipsychotics, a meta‐analysis of observational studies was performed. Adjusted odds ratio (OR) of SCD/SUD with 95% confidence intervals (CI) were extracted and pooled; heterogeneity was studied using Q statistic and I2 index, and its potential causes (e.g., hERG blockade potency) explored using meta‐regression. Two cohort (740,306 person‐years) and four case–control (2,557 cases; 17,670 controls) studies, investigating nine antipsychotics, were included. Compared with nonusers, the risk was increased for quetiapine (OR = 1.72, 95% CI: 1.33–2.23), olanzapine (OR = 2.04, 1.52–2.74), risperidone (OR = 3.04, 2.39–3.86), haloperidol (OR = 2.97, 1.59–5.54), clozapine (OR = 3.67, 1.94–6.94), and thioridazine (OR = 4.58, 2.09–10.05). Heterogeneity was found (Q = 20.0, P = 0.01; I2 = 60.0%), and the increasing mean hERG blockade potency (P = 0.01) accounted for 43% of this. The SCD/SUD risk differed between individual antipsychotics, and mean hERG blockade potency could be an explanatory factor. This should be considered when initiating antipsychotic treatment.

Some antipsychotic drugs have been shown to increase the risk of sudden cardiac death (SCD) and/or sudden unexpected death (SUD). 1,2 One potential mechanism is via blockade of the cardiac potassium channel coded by hERG, the human Ether-a-go-gorelated gene, usually referred to as the hERG channel, 3 which is responsible for the rapid delayed rectifier potassium current (I Kr ). This blockade can result in delayed cardiac repolarization, prolongation of the QT interval on the ECG, 4 and an increased risk of the polymorphic ventricular tachycardia termed torsade de pointe, a cardiac rhythm disorder that may result in cardiac arrest.
Pharmacoepidemiological studies have previously been performed to investigate the risk of SCD/SUD associated with the use of antipsychotic drugs. Some of these compared typical and atypical antipsychotics and found no interclass difference, 1,5,6 but all had limited power to compare risks associated with the use of individual antipsychotics. 1,6 Thus, a meta-analysis of observational studies was conducted to evaluate the risk of SCD/SUD associated with the use of individual antipsychotic drugs and to explore sources of potential heterogeneity among drugs. 40 references underwent full-text examination (see Supplementary Information for selection process details), which generated the final selection of six observational studies eligible for the meta-analysis (Supplementary Figure 1). 1,2,[6][7][8][9] Characteristics of included studies Of the selected six studies, two were cohort studies 1,6 and four were case-control studies. 2,[7][8][9] Four studies used medical record databases (four used primary care records, 2,6,9 one used hospital records 8 ), one used claims database records, 1 and one was a field study. 7 Four studies provided risk estimates for SCD, 1,6,8,9 and two for SUD. 2,7 Relative effect estimates were reported as odds ratio (OR), relative risk (or risk ratio, RR), or IRR (incidence rate ratio; Table 1).
Concerning quality, 10 the two cohort studies obtained the highest possible score on the Newcastle-Ottawa Scale (nine stars, see Supplementary Table 1). For the case-control studies, this score ranged from seven to nine stars (see Supplementary Table 2); weaknesses concerned the selection of controls from hospital rather than community settings, 7,8 and lack of information on response rate. 8 Risk estimates were available for nine individual antipsychotics. Haloperidol was investigated in five studies; chlorpromazine, quetiapine, risperidone, and thioridazine were studied in three studies each; clozapine, flupentixol, fluphenazine, and olanzapine in one study each ( Table 1). In the cohort studies, 435 cases of SCD/SUD were found in current users of antipsychotics (cumulated follow-up: 115,921 person-years), and 993 in the nonuser group (648,414 person-years); in case-control studies, 132 exposed cases (2,425 nonexposed cases) and 169 exposed controls (17,501 nonexposed controls) were found.

Meta-analysis
Risk estimates of SUD/SCD for individual drugs ranged from OR 0.06 (95% CI: 0.00-6.00) for fluphenazine, to OR 9.40 (95% CI: 0.21-420.75) for flupentixol, for which only one estimate was available from the selected studies ( Figure 1). The risk was nonsignificantly increased for chlorpromazine use compared to nonuse (OR 1.66, 95% CI: 0.83-3.29; three studies). It was found to be significantly increased for quetiapine ( 1 ; the attributable number of cases was then calculated for the studied antipsychotics, and it varied from 1.0 per 1,000 person-years (95% CI: 1.3-1.5) for quetiapine, to 5.1 per 1,000 person-years (95% CI: 1.5-13.8) for thioridazine ( Table 2).
The R 2 index, used to quantify the proportion of the variability in the risk with each potential source of heterogeneity, 14,15 for mean hERG blockade potency was 43%, while for other covariates were nil or very low (2% for study design).

DISCUSSION
The present meta-analysis included data for nine antipsychotics, the use of six of which has been associated with an increased risk of SCD/SUD. Heterogeneity was found between drugs; the highest risk was associated with clozapine, haloperidol, risperidone, and thioridazine, and the data indicate that there was a considerable number of attributable fatal cases related to these drugs. Meta-regression indicated that the risk of SCD/SUD increased with increasing of mean hERG blockade potency.
hERG channel blockade is recognized as one of the mechanisms for QT prolongation, which is an intermediate finding for torsade de pointe and ventricular arrhythmia. 3 Although this association is evident, up to today the hERG affinity has not been related to SCD/SUD risk in clinical practice. The present study supports the hypothesis that hERG affinity of antipsychotics is an independent risk factor for SCD/SUD. In the present analysis, R 2 index indicates that hERG blockade potency explains more than 40% of the heterogeneity found among drug estimates. However, the performance of the R 2 index in meta-regressions is limited when fewer than 40 estimates are included in the analysis. 15 Moreover, published data concerning hERG blockade potency of antipsychotics is limited, and other mechanisms may also be involved (e.g., drug effects on other myocardial ion channels). Despite these considerations, it is of note that among the variables we tested as a potential cause of heterogeneity, mean hERG blockade potency was the only one that was significantly associated both in univariate and multivariate analyses; forcing all variables in the model did not alter this result (data not shown).
The high risk of cardiac arrhythmias associated with thioridazine use is widely accepted, resulting in the drug being withdrawn in 2005 in several countries; this study confirmed its strong association with SCD/SUD. The increased risk found of SCD/SUD for clozapine should be considered with caution: the risk estimate was not precise (only one study for this drug), and the reported hERG blockade potency is highly variable between assays (50% inhibitory concentration [IC 50 ] from 0.32 to 2.63 mmol). [16][17][18] More importantly, as a second-line treatment for resistant Studied drugs (n exposed cases; n exposed controls) Jolly et al.  20 which, if unrecognized in fatal cases, could be classified as a sudden death. The risk associated with haloperidol was also high. This is consistent with its extremely high potency for hERG channel blockade, and its myocardial cell concentration being higher than other antipsychotics. 21 Nevertheless, the interpretation of this result requires caution due to the heterogeneity found between its estimates, with consequently wide 95% CIs. This could be related to multiple indications of haloperidol (including delirium or nonpsychiatric conditions), and different routes of administration. Further investigations are required to clarify this issue and to make more definitive conclusions about its risk.
For risperidone, the magnitude of risk was consistent between individual study estimates. Risperidone has one of the highest mean hERG channel blockade potency, and a high intramyocardial concentration. 21 A potential alternative explanation is that risperidone is preferentially prescribed to at-risk patients, such as the elderly with dementia, for which is it indicated. However, there was a lack of data regarding the population using risperidone that limited investigation of such aspects, as was the case for other drugs considered herein.
Quetiapine, a drug with low mean hERG blockade potency, showed a lower increase in the risk of SCD/SUD, and estimates among the three included studies were consistent. This study indicates that quetiapine is the atypical antipsychotic related to a lower increased risk of SCD/SUD.
Olanzapine, which also has low mean hERG channel blockade potency, was also associated with more limited increase in risk; although only one study was retrieved, this was large. Other epidemiological evidence indicates that olanzapine has a lower risk of SCD when compared with haloperidol, but that this risk is higher than that for quetiapine. 22 For chlorpromazine and flupentixol, the increase in risk was not significant; there was no evidence of heterogeneity among estimates for chlorpromazine, and for flupentixol only one study was included, and not many patients were exposed, limiting the conclusions that can be drawn. According to our hypothesis concerning hERG channel blockade, the high potency of flupentixol, however, does suggest the need for further investigations. Similarly, the estimate for fluphenazine was also derived from a single case-control study, and despite being associated with a point estimate less than 1.0, the 95% CIs are too wide to allow any meaningful conclusion. This indicates a clear knowledge gap concerning the arrhythmogenecity of these two drugs, yet observational studies needed to fill this gap may be difficult to perform due to the ever-decreasing use of such drugs (flupentixol is not even marketed in the US, for instance).
This meta-analysis has certain limitations. For instance, studies evaluating SCD with those evaluating SUD, which could also be related to other diagnoses, such as tonic-clonic epilepsy, were pooled. Nevertheless, the included studies that did investigate SUD, alone or in relation with SCD, used strict criteria to avoid the inclusion of unexpected death from causes other than arrhythmias. 2,7,23 Despite this wide definition, the present analysis was only able to pool data for nine of the more than 50 antipsychotics available worldwide, but these do represent the most commonly used typical and atypical antipsychotics. Also, the analysis according to dose was performed in only two of the included studies, 1,7 and no information was found concerning different formulations of included drugs (e.g., prolonged release) or different routes of administration. This advocates for the conduct of additional observational studies focusing on the risk of SCD/SUD associated with the use of individual antipsychotics, such as those envisioned in the ARITMO project. Another important aspect is that patients with psychiatric bipolar disorders, schizophrenia, and dementia have an increased risk of all-cause mortality in comparison with the general population; thus, an indication bias cannot be excluded. It is likely that this would not be differential across the drugs used as first-line treatments, while it is likely to have been greater for clozapine, a second-line treatment reserved for more severe psychoses, and for atypical antipsychotics, which are more commonly used in the elderly. Studies did address this issue by matching or adjusting for psychiatric disorders, cardiovascular diseases, or through propensity scores; in addition, one study also adjusted for hypokalemia and alcohol abuse, 2 and a second for use of drugs that could induce hypokalemia. 9 The results provided by the latter studies did not differ from those provided by others. Residual confounding could also be present in all included studies that, even matched or adjusted, remain observational and thus cannot control for all potential risk factors of SCD/SUD. One of these potential residual confounders is obesity, which was considered in none of the included studies. Obesity was recently related to SCD in patients in middle-aged, nonsmoking individuals 24 ; nevertheless, this increased risk seems mediated by traditional cardiovascular risk factors (such as arteriosclerosis), which were taken into account in all of the included studies.
The results may also be affected by the quality of the included studies. According to the Newcastle-Ottawa scale, the quality of five studies reached the highest value; for others, it was one to two points below this. The quality of the studies reported by Kenbubpha and Silpakit 7 and by Reilly et al. 8 was adversely affected by the selection of controls in the hospital as opposed to community settings, but, as in these case-control studies cases were selected from the same population, this may not be too great a concern. The study reported by Reilly et al. 8 was also affected by nonresponse rate, as for certain variables there were missing data, but it was not made clear in the published article whether there was any imbalance across groups.
Heterogeneity is also a concern in meta-analyses, yet it is common when meta-analyses of observational studies are performed, 25 and it may reach very high values. 26,27 When heterogeneity is found intradrug this could be considered a limitation, but when this is interdrug, this could be considered as an interesting aspect to investigate the risk profiles of drugs in a real-life setting. In this study we investigated heterogeneity using meta-regression, which could also be affected by certain limitations, such as aggregation or ecological bias. 28,29 However, among tested variables, aggregation bias could affect only age and gender, while the multilevel metaregression performed excluded ecological bias. The estimations for each antipsychotic are affected by a certain degree of imprecision. This affects in particular drugs rarely used in clinical practice, such as flupentixol or fluphenazine. From a general point of view, imprecision leads to an underestimation of heterogeneity among estimates; thus, in the present study it could have reduced the chance of finding a difference in risk among studied drugs.
More generally, a publication bias could have affected the results. Nevertheless, in observational studies this is difficult to detect and to evaluate, as registers are not systematically used, as they are for clinical trials. In addition, due to the differential risk among the investigated drugs, a funnel plot could be appropriate and informative when at least 10 estimates for the same drug are available, 30 which was not the case in the present meta-analysis. The existence of a publication bias could be considered unlikely for different reasons: the first studies were published in 2002; 7,8 only two of the six studies were privately funded; 2,6 the included articles studied antipsychotics as a class, without any specific analysis for a single drug; this meta-analysis evaluated more than  700,000 person-years for cohort studies, and more than 3,000 cases for case control ones; the results are coherent with pharmacological profiles of the studied drugs. If a selective publication for positive results could be expected for observational studies, the large effect size found in this meta-analysis strongly reduces the possibility that new evidence substantially changes the results of the present meta-analysis.
In conclusion, this study strongly suggests that the risk of SCD/SUD differs between individual antipsychotics, and that this difference in absolute terms is clinically important. It also suggests that mean hERG blockade potency is an explanatory factor for this difference in real-life situations. This should be considered when initiating antipsychotic treatment, in particular as indications for these have been widened to nonschizophrenic patients in whom arrhythmogenic risk is less extensively studied.

Data sources and data extraction
As part of the ARITMO project, an extensive systematic review was conducted to identify observational studies that evaluated the risk of ECG disorders and ventricular arrhythmias related to antipsychotics, antiinfectives, and antihistamines. The results concerning use of individual antipsychotics and risk of SCD/SUD in comparison with nonuse are presented herein. The studies evaluating SCD, defined as sudden natural death attributable to cardiac causes or nonattributable to a noncardiac cause, and SUD, defined as sudden death both of cardiac origin or other natural origin, were considered eligible to be included in the present meta-analysis.
Studies eligible for inclusion in this meta-analysis were cohort studies (prospective or retrospective), case-control studies, or case-based studies including case-crossover and self-controlled case series, providing adjusted risk estimates of SCD/SUD. For this study, a protocol specifying the meta-analysis objective and context, the principles and modalities of the literature search, and the data analysis was developed during the ARITMO project period. This study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) statement, and Meta-analysis of Observational Studies in Epidemiology (MOOSE) reporting guidelines (see Supplement File). 31,32 References meeting the inclusion criteria were searched in Medline, EMBASE, and ISI Web of Knowledge. The search strategy with detailed literature keywords used in the Medline database is reported in the data supplement. EndNote X4 for Macintosh (Thomson Reuters) was used to compile the bibliography. This search was last updated on February 13, 2014. The process used for the identification and selection of the retrieved studies is detailed in the data supplement.
Data were extracted from the selected studies using a standardized form. These included the definition of the events of interest, the number of cases in the exposed and nonexposed group for cohort studies, the number of exposed cases and exposed controls for case-control studies, and the adjusted risk measure of SCD/SUD for each individual antipsychotic.
The Newcastle-Ottawa Scale was used for assessing the quality of the included studies. Each study was scored from zero to nine stars for the selection and the comparability of the groups, and for the ascertainment of either the exposure for case-control or outcome of interest for cohort studies. 10

Meta-analysis
The overall risk of SCD/SUD associated with the use of each individual antipsychotic was estimated. Adjusted estimates of risk were extracted from each selected study, and included in a forest plot as OR, which is a good risk measure when the diseases are rare, and when different study designs are pooled in a single meta-analysis. For each antipsychotic, the pooled OR with 95% CI for use vs. nonuse was computed using the inverse variance method and random-effect models. 11 Statistical heterogeneity was evaluated using the Q statistic (with P < 0.10 considered significant) and the I 2 index 33 between the individual risk estimates for a given drug (intradrug heterogeneity), and between the pooled estimates of each drug (interdrug heterogeneity). The meta-analysis was conducted using Review Manager software (RevMan v. 5.2, The Nordic Cochrane Centre, Cochrane Collaboration); all P values were two-sided.
The cohort study reporting the lowest incidence rate of SCD/SUD in the nonuser group was used to calculate the number of attributable cases (and the corresponding 95% CI) per 1,000 person-years. The incidence rate in this nonuser group was multiplied by the pooled OR minus 1.0 for each antipsychotic associated with an increased risk of SCD/SUD.

Meta-regression analysis
Random-effect meta-regression analyses were performed to explore sources of heterogeneity between the individual estimates of SCD/SUD risk. 11,12 Potential sources of heterogeneity considered included: individual study  Table 3). [16][17][18][34][35][36][37][38][39][40][41][42][43][44][45][46] The ARITMO consortium researchers retrieved the references used for this calculation via a systematic review of preclinical data of hERG channel blockade potency. 47 Each potential source of heterogeneity was investigated using univariate meta-regression models. A multivariate meta-regression model was then built using manual backward elimination including only sources associated with P < 0.05. 13 The R 2 index was used to quantify the proportion of the variability in the SCD/SUD risk associated with each potential source of heterogeneity. 14,15 The R 2 index informs the practical significance, or the degree of influence, of a variable in the heterogeneity of the effect sizes in a meta-analysis (e.g., if a variable with R 2 5 20%, then this variable explains 20% of the heterogeneity). The meta-regression analyses were performed using R software (v. 3.0.3) via the "metafor" package. 48 Additional Supporting Information may be found in the online version of this article.

ACKNOWLEDGMENTS
This study is part of a research project that has received funding from the European Community's Seventh Framework Program under grant agreement number 241679: the ARITMO project. The corresponding author confirms the full access to all data in the study and the final responsibility.