Antidepressant use and risk of epilepsy and seizures in people aged 20 to 64 years: cohort study using a primary care database

Study cohort

Patient data from 687 general practices across the UK was extracted from the QResearch primary care database (version 34), a large repository of anonymised health records for more than twelve million patients registered with general practices using the Egton Medical Information System (EMIS).

Patients were selected for the cohort if they had a first recorded diagnosis of depression between the ages of 20 to 64, and within the study period (1/1/2000 to 31/7/2011). To allow for at least 12 months of follow-up the study end date was set to 1/8/2012. Identification of patients with a diagnosis of depression was achieved using clinical Read codes that have been used in previous studies [4, 11]. Depression in the UK is largely diagnosed and treated in primary care without the input of specialist assessment and there is no widespread formal screening for depression using validated tools.

The study entry date for each patient was defined as either: the date of their first recorded diagnosis of depression, or the date of their first antidepressant prescription if that occurred before the diagnosis date and within 36 months of diagnosis. To ensure only patients with a first diagnosis of depression were included in the cohort, eligible patients had to be diagnosed with depression at least 12 months after registering with the practice and after the date the EMIS system was installed.

Patients with a recorded diagnosis of depression before the study entry date and those prescribed antidepressants before this date, before their date of registration with the practice, before the age of 20 or more than 36 months before their first recorded diagnosis of depression were excluded. Temporary residents, patients with diagnoses of schizophrenia, bipolar disorder or other types of psychosis, or those prescribed lithium or anti-manic drugs at the study entry date were also excluded. In addition, although we have included the number of patients receiving MAOI prescriptions, and the number with a previous history of epilepsy/seizures in descriptive results, before proceeding with the main analyses, patients satisfying either of these criteria were dropped from the cohort.

Follow-up continued until the earliest date of: the patient’s first recorded diagnosis of epilepsy/seizures, the patient left the practice, death, or the end of the follow-up period (1/8/2012). This became the patient’s study exit date.

Outcomes

The outcome was the first recorded diagnosis of epilepsy or a seizure in the patients’ primary care medical record. This outcome was identified using 88 separate Read codes for epilepsy/seizures (see Additional file 1: Table S1).

Exposure data

Data on antidepressant prescriptions during follow-up were extracted, including the date of each prescription, the antidepressant drug name, the dose of each tablet in milligrams, dosage instructions, plus the number of tablets or capsules prescribed. The duration in days for each prescription was calculated by dividing the number of tablets prescribed by the number to be taken per day indicated by the dosage instructions. Where dosage directions were missing or unclear (just under 5 % of prescriptions) we used the median duration from prescriptions for the same drug where dosage was recorded, accounting for the number of tablets prescribed, as in our previous study [4, 9]. The total daily dose was generated by multiplying the dose per tablet by the number of tablets to be taken each day. In order to make direct comparisons between different antidepressant classes we converted the total daily dose to a defined daily dose (DDD) using the assumed average daily maintenance dose as given for each antidepressant on the WHO website (www.whocc.no/atc_ddd_index/). Where prescriptions for the same drug were issued on the same day we counted them as a single prescription and summed the doses.

We classified antidepressant drugs according to the classes in the British National Formulary (BNF) which gives four main categories: tricyclic and related antidepressants (TCAs - section 4.3.1), selective serotonin reuptake inhibitors (SSRIs - section 4.3.3), monoamine oxidase inhibitors (MAOIs - section 4.3.2), and other antidepressants (section 4.3.4). We defined an additional fifth category as combined prescriptions, comprising prescriptions for different drugs, either within a class or from different classes, given on the same date; these prescriptions were not included in analyses of dose.

In a similar way to the previous study [4, 9] the 11 most frequently prescribed individual antidepressants were identified.

Confounding variables

Both demographic and baseline characteristics and comorbidities were included as confounding factors in the adjusted analyses. These were a general pool of confounders considered likely to be associated with either the selection of an antidepressant or the risk of the outcome as in the previous study [9].

The demographic features were: age at study entry, gender, year of index diagnosis of depression, severity of index diagnosis of depression (categorised as mild, moderate or severe, using the Read code classification published by Martinez and colleagues [11] and some additional classification by a member of the study team (RM) of further Read codes for depression, present in our data but not included in the Martinez study, see Additional file 1: Table S2), Townsend deprivation score (in fifths) [12], smoking status (non-smoker, ex-smoker, current smoker coded as light (1–9 cigarettes/day), moderate (10–19 cigarettes/day) or heavy (≥20 cigarettes/day)), alcohol intake (none, trivial (less than 1 unit per day), light (1 to 2 units per day), medium (3 to 6 units per day), heavy (7 to 9 units per day) and very heavy (more than 9 units per day)), and ethnicity categorised as either white/not recorded or non-white (comprising Indian, Pakistani, Bangladeshi, other Asian, black African, black Caribbean, Chinese and other/mixed groups).

Statistical analysis

We used Cox’s proportional hazards models to analyse the time to the first recorded diagnosis of epilepsy/seizures and to assess any differences in risk associated with the antidepressant prescribed. These analyses used robust standard errors to account for clustering of patients within general practices. Patients with a previous history of epilepsy/seizures were dropped from the cohort before proceeding with these analyses. Our analysis treated antidepressant use as a time-varying exposure which partitioned follow-up time for each patient into treatment periods by antidepressant class and also any periods of non-treatment between prescriptions. We regarded a patient as not exposed to treatment during periods where there was a gap of 90 days or more between the end of one prescription and the start of the next. We based our main results on the first 5 years of follow-up after study entry, and also conducted analyses for the first year of follow-up and for the whole of follow-up time. We calculated unadjusted and adjusted hazard ratios with 95 % confidence intervals, adjusting for the confounding variables listed above.

The analysis of antidepressant treatment was conducted in five main ways: by drug class (with no current treatment as the reference group), drug class (with SSRIs as the reference group), dosage, the 11 most common antidepressants, and by time since starting and stopping antidepressant treatment. Table 3 shows the categories used. For the dosage analysis we did not calculate DDDs for combined prescriptions, so only used one category for them. For the time since starting and stopping analysis, periods of non-treatment from 183 days after stopping treatment were coded as no use of antidepressants.

We tested for interactions between antidepressant drug class and age, sex and ethnicity.

We determined statistical significance using a two-tailed P value of <0.01, and used Wald’s significance tests to assess differences between drug classes. We used log-log plots and tests of the proportional hazards assumption via Schoenfeld residuals, to check the validity of our models.

Numbers needed to harm (NNH) and absolute risks were calculated over 1 and 5 years based on the method described for numbers needed to treat (NNT) by Altman et al. [13, 14] using adjusted hazard ratios. The NNH for a particular drug gives the number of patients that would need to be treated with that drug, over a certain period of time, for there to be one extra case of epilepsy/seizures more than that occurring in the untreated, comparison group. The quantity is generated by first calculating the absolute risk increase by subtracting the event rate in the treated group from that in the untreated control group (at a specific time), and then taking the reciprocal of that number. The NNH assumes a causal relationship between each drug and the risk of epilepsy/seizures.

All analyses were conducted with Stata/MP v12.1.

Sensitivity analyses

We performed two sensitivity analyses. In the first we repeated the main analyses, but excluded patients who received no antidepressant treatment during follow-up. The reason for this analysis was to examine whether any differences between the untreated and treated patients that we were not able to account for in the analysis, such as a preference for psychotherapy rather than drug treatment, could introduce a form of selection bias into our results.

In the second we repeated the analyses excluding patients on anticonvulsants at baseline, in case these had a previous history of epilepsy that was unrecorded and could bias the results, although anticonvulsant use is not necessarily an indication of epilepsy or seizures as these drugs are sometimes used in the treatment of other conditions such as neuropathic pain [15].

Approvals

The project has been independently peer reviewed and accepted by the QResearch Scientific board and has been approved in accordance with the agreed procedure with the Trent Research Ethics Committee (reference No MREC/03/4/021).

Demographic characteristics of the cohort

Table 1 shows the baseline characteristics of the cohort. There were 146,028 (61.1 %) females and 92,935 (38.9 %) males. The mean age was 39.5 years (SD = 11.1).

The most common comorbidity at baseline was asthma and/or chronic obstructive airways disease (31,816 patients, 13.3 %), followed by hypertension (17,217 patients, 7.2 %). The most common medication at baseline was antihypertensives (25,344 patients, 10.6 %). There were 2672 (1.1 %) patients on anticonvulsants at baseline.

Antidepressant treatment during follow-up

A total of 209,476 (87.6 %) patients received antidepressant treatment during follow-up, with 3,337,336 antidepressant prescriptions issued during the whole of follow-up. This mainly comprised SSRI prescriptions (2,379,668, 71.3 %), with 533,798 (16.0 %) prescriptions for TCAs, 422,079 (12.7 %) for other types of antidepressant and only 1791 (<0.1 %) prescriptions for MAOIs. The median duration of treatment during follow-up was 221 days (interquartile range 79 to 590 days), with 5.5 % of patients having a total duration of treatment of 5 or more years. Within the whole study period there were over 1.3 million person-years of follow-up.

See Additional file 1: Table S3 for a breakdown of baseline characteristics for the cohort by the type of antidepressant first prescribed, including a column for those patients not treated with antidepressants during follow-up.

Number of cases of epilepsy/seizures at baseline and during follow-up

At baseline 3325 patients (1.4 %) had a previous diagnosis of epilepsy and/or seizures. We excluded these patients and also, due to small numbers, those with MAOI prescriptions (n = 156) from further analyses leaving 235,489 patients (72 % of the initial sample).

During the entire follow-up period 1126 patients had a first diagnosis of epilepsy/seizures. Within the first 5 years of follow-up 878 patients (416 women and 462 men) had a first diagnosis of epilepsy/seizures, giving an incidence rate of 99 per 100,000 person-years (76 in women and 135 in men).

Associations with epilepsy/ seizures over 5 years of follow-up

All dose categories other than low doses of SSRIs and other antidepressants were associated with a significantly increased risk (at P < 0.01) of epilepsy/seizures compared to periods of no treatment. Hazard ratios increased with dose for SSRIs but for TCAs and other antidepressants the middle dose categories had the highest hazard ratios.

There were significantly increased hazard ratios for 8 of the 11 most commonly prescribed drugs with the highest risks associated with trazodone (adjusted HR = 5.41, 95 % CI 3.05 to 9.61), lofepramine (adjusted HR = 3.09, 95 % CI 1.73 to 5.50), and venlafaxine (adjusted HR = 2.84, 95 % CI 1.97 to 4.08). Escitalopram, sertraline and mirtazapine, were not associated with a significant increase in risk (at P < 0.01). Wald’s tests however did not indicate any significant difference overall between all the individual antidepressant drugs (P = 0.029).

When compared against the drug citalopram, only trazodone (HR = 2.66, 95 % CI 1.49 to 4.76) had a significantly increased risk of epilepsy/seizures. Full results of analyses of direct comparisons with mid-dose SSRIs and citalopram are shown in Additional file 1: Table S4.

In the analysis of duration (Table 3 and Fig. 1), there were no clear patterns of risk and all hazard ratios were significant except for SSRIs in the first 28 days of treatment and other antidepressants in the 29–84 days category. In the time since stopping treatment categories only 3 were statistically significant: the first 28 day period for SSRIs (adjusted HR 2.28, 95 % CI 1.50 to 4.36) and the 29–84 day periods for TCAs (adjusted HR 2.79, 95 % CI 1.65 to 4.72), and other antidepressants (adjusted HR 4.46, 95 % CI 2.46 to 8.09). However numbers were small in some categories.

Interaction models

There were no statistically significant interactions between drug class and either age, sex or ethnicity.

Sensitivity analyses

In the first sensitivity analysis excluding untreated patients, although all hazard ratios increased slightly, only one result, for low dose SSRI, went from borderline significance to statistical significance (HR from 1.80 to 1.88, with P moving from 0.011 to 0.006) (Additional file 1: Table S5).

In the second sensitivity analysis omitting patients on anticonvulsants at baseline (Additional file 1: Table S6), most hazard ratios increased with the significance levels largely remaining unchanged. For individual drugs the only hazard ratio to decrease was for escitalopram, going from 1.49 to 1.33. The greatest increase was for trazodone (HR from 5.41 to 5.95).

Associations with epilepsy/seizures over 1 year’s follow-up

We repeated the analyses restricted to the first year of follow-up and the hazard ratios tended to be lower and with wider confidence intervals due to the reduced number of events (Additional file 1: Tables S7 and S8).

Absolute risks of epilepsy/seizures

The drug class with the highest NNH is the SSRIs: 2014 patients would have to be treated for a year for there to be one extra case of epilepsy/seizures. Over five years 312 patients would have to be treated for an extra case. At the other end of the spectrum, only 166 patients would have to be treated with a combined prescription for five years, for there to be an extra case. For individual drugs the lowest NNH is for trazodone, where for one extra case of epilepsy/seizures to occur 602 patients would have to be treated for one year, and only 65 patients would have to be treated for five years. Although the NNH over one year for lofepramine and venlafaxine are higher at 2565 and 749 respectively, they are only 138 and 156 for five years of follow-up. Of all the individual antidepressants, these three drugs have the lowest NNH over five years of follow-up, with trazodone having almost twice the absolute risk as the other two.

Comparison with other studies

Whilst the association between epilepsy/seizures and depression [8, 16, 17] and that between antidepressant use and epilepsy/seizures [18] has been previously documented, some of this evidence is conflicting. For example some studies claim the effects of antidepressants are anticonvulsant rather than pro-convulsant [19] and other studies claim that most if not all antidepressants lower the seizure threshold [3].

There are studies concerned only with a population already predisposed to the outcome [20–24], however their results may not be directly comparable to ours as we omitted patients with a history of epilepsy/seizures. One of these studies [20] concluded that patients could be safely treated with antidepressants from the TCA or SSRI groups without any significant increase in seizure frequency. However the study only looked at the first 12 months of treatment, only included two of the antidepressants that we did (amitriptyline and paroxetine) and was much smaller than our study (n = 421).

Another large study utilising drug safety data [24] concluded that SSRIs and the antidepressants mirtazapine, venlafaxine and duloxetine might be more appropriate for treating depression than TCAs, in patients with an enhanced seizure risk. They based their analysis on counting the number of grand mal seizures in patients undergoing antidepressant treatment in psychiatric inpatient settings. A limitation of this analysis is that seizures were probably underreported.

A further issue is whether the antidepressants are given in low doses, normal therapeutic doses, or taken in overdose. For example, there is evidence that citalopram has the ability to induce seizures if taken in overdose [25] and that it is more toxic and more likely to provoke seizures in overdose than escitalopram [26]. However a review by Nemeroff [27] concluded that doses of between 20–60 mg of citalopram do not increase risk of seizure.

A study utilising Food and Drug Administration (FDA) Summary Basis of Approval (SBA) Reports in the United States [28] concluded that many second-generation antidepressants, several of which are included in our study (including citalopram and venlafaxine), have anticonvulsant properties. They based their conclusions on comparing seizure incidence between treatment and placebo arms of clinical trials, with numbers of participants less than 5000 in most instances, and with short follow-up periods (the mean duration of the treatment arms was only 116 days, with the mean duration of the placebo arms only 78 days). They also found an incidence of seizures in the placebo groups approximately 19 times that seen in the general population.

These results may seem to contradict our own, but many of the placebo arms of the studies included in the Alper [28] paper did not report the number of seizures, a limitation which they acknowledge. In addition, although many of the antidepressants they looked at are also included in our study, they did not treat each antidepressant individually, but rather they combined them together and compared the overall seizure rate with the placebo arms. Also, seizures are only recorded in SBA reports if the FDA staff physician deems the event was unlikely to have been provoked by an external factor, such as alcohol withdrawal, and could be safely attributed to the drug in question. Our study included all events, without an attempt to distinguish between the possible causes. Generally clinical trials are designed to answer questions regarding benefits of treatments rather than safety and full reporting of adverse effects is often unavailable [29].

In an analysis [30] of the WHO adverse drug reactions (ADR) database, the authors concluded that antidepressant treatment, lowers the seizure threshold and provokes seizures. They looked at 9 of the 11 most commonly prescribed antidepressant medications that we did. There were some differences in their results compared to ours though, since they showed trazodone to have one of the lowest ratios of suspected seizures. There were many limitations to their study however, including the fact that the total number of patients treated with a particular drug was unknown making direct comparisons with our results inappropriate.

In our previous study in people aged 65 years and over [4, 9] we found that SSRIs and other antidepressants were associated with an increased risk of epilepsy/seizures. This situation has been repeated in the present study, with adjusted hazard ratios very similar to the earlier study. Also, citalopram, paroxetine, and venlafaxine were found to be significantly associated with an increased risk of epilepsy/seizures in both studies.

The significant association with sertraline in the earlier study is no longer present whereas there is now a significant association with fluoxetine. But the main difference is that in the present study we have found a significant association with TCAs when before there was none.

The previous study reported an adjusted hazard ratio of 1.02 for TCAs for older people but for the present study it was 2.32. In the previous study, none of the four TCAs were significantly associated with epilepsy/seizure risk, however in the present study they are all associated with a statistically significant increased risk with the highest hazard ratio being for trazodone.

Why TCAs should be significantly associated with an increased risk of epilepsy/seizures in the younger sample but not the older sample is unclear, particularly since the proportion of TCA prescriptions in the present study is far less than the earlier study: 31.6 % vs. 13.3 %. Although there were fewer low dose TCA prescriptions in the new study compared to the older study (60 % vs. 70 % were for ≤ 0.5 DDD), this is not enough to explain the difference between the two results.

Strengths and limitations

We know of no other study utilising a sample as large, and followed-up for as many years as our cohort. We were able to analyse a range of antidepressant medications, including individual drugs and accounted for many confounding factors. We also calculated absolute risks that enable comparisons to be made between different drugs that can be explained to patients.

We have used a very large sample representative of primary care populations where most treatment decisions are made, so our results are likely to be generalizable to people aged 20 to 64 years with a diagnosis of depression in primary care.

Our study included all eligible patients within the study time period, and did not exclude people with any comorbid disease or taking any medications other than those explicitly stated in the exclusion criteria (patients with diagnoses of schizophrenia, bipolar disorder or other types of psychosis, and those prescribed lithium or anti-manic drugs). This limited the potential for selection bias which can affect observational studies. Recall bias was avoided as data on prescriptions and confounding factors were recorded prospectively on the computer system.

We had access to detailed information on antidepressant prescriptions, including class of antidepressant, drug name, dosage and duration, so were able to investigate associations with these factors.

To reduce the effects of bias we adjusted for many confounding factors, however we were unable to control for certain factors such as family history of seizures, seizure history without diagnosis, neurological abnormalities, or other factors known to affect seizure threshold such as antibiotic use [31–34] so there may be some residual confounding. Although we adjusted for depression severity, this was based on a logical but crude classification of diagnostic Read codes that has not been formally validated, since depression severity scores are not routinely recorded in general practice.

We excluded patients with a history of epilepsy/seizures at baseline thereby reducing the possibility of a drug being prescribed on the basis of whether it was known to affect the seizure threshold (channelling bias).

The generally similar patterns of risk for the different drug classes and individual drugs might indicate that depression itself rather than its treatment could be the cause of the increased rates of epilepsy/seizures. We restricted our cohort to patients who had a diagnosis of depression to reduce this potential indication bias however we compared rates during periods of treatment, with untreated periods when the depression may have resolved. Several studies have reported that depression is associated with increased rates of epilepsy/seizures including a bidirectional relationship [8, 16, 17], although it is difficult to fully separate the effects of depression from those of antidepressant treatment.

We do not know whether patients took their medication at the exact prescribed dose or whether they completed the course of medication. As a result, this may mean some periods of antidepressant treatment have been misclassified.

Our outcome was a diagnosis of epilepsy/seizures as recorded in the general practitioners record. The nature of the database and the study design meant that this could not be formally adjudicated as might be the case in a clinical trial. However since epilepsy is considered a major diagnosis with consequences which can affect a person’s employment or ability to drive, we think it is unlikely that it would be entered in the record without there being sufficient evidence to support the diagnosis. Also epilepsy has been a condition included in the GP Quality and Outcomes framework over recent years and this is likely to have resulted in increased diligence regarding the accuracy of the diagnosis recorded on the computer system. Another limitation of our study is that we were not able to distinguish between seizures that were epileptic and non-epileptic in nature, such as psychogenic non-epileptic seizures. There is no reason to believe however that different antidepressants would have a differential effect on inducing non-epileptic seizures.

Implications of our findings

Whilst we acknowledge our results need further confirmation via future studies, there are possible implications of our findings for the treatment of depression.

We have found that treatment with all antidepressants is associated with an increased risk of epilepsy/seizures, with some types of antidepressant being associated with a much higher risk than others. This suggests there is a need to inform patients and their doctors about the increased risks of continued antidepressant therapy that accumulate over long periods of treatment. Consideration of each patient’s individual circumstances must be taken into account, allowing discussion between the clinician and the patient, enabling a shared decision to be reached on future treatments.

Due to the increased absolute risks of epilepsy/seizures that occur under long term use, there is a need for an individual clinical risk-benefit assessment for any patient facing antidepressant treatment over a period of five or more years, bearing in mind the absence of evidence supporting longer term treatment with antidepressants. This may be of particular importance to people with mild and infrequently recurring depression, or those with additional risk factors for seizures such as evidence of cerebrovascular disease.

In instances of mild depression in particular there may be a case for stopping treatment, or a reduced daily dosage, or switching to a different antidepressant that carries a lower risk or seeking alternative treatments, such as cognitive behavioural therapy, rather than continuing with long-term antidepressant treatment which would carry a further increased risk of epilepsy/seizures.

In instances where the choice of antidepressant and dose has already been driven by factors such as how well the patient tolerates the drug or the severity of the depression, our findings can still inform the continued treatment of the patient as the accumulating risk that occurs over time that we have demonstrated can be communicated to the patient by the clinician. An informed decision can then be reached regarding continued antidepressant treatment or alternative options from that point forwards.