Prediction of risk of seizure recurrence after a single seizure and early epilepsy: further results from the MESS trial

BACKGROUND
The MRC Multicentre trial for Early Epilepsy and Single Seizures (MESS) showed a reduced risk of further seizures in patients, for whom treatment with antiepileptic drugs was uncertain, who were randomly assigned immediate treatment compared with delayed treatment. However, there was no evidence of an effect on [corrected] long-term remission rates. This study was undertaken to assess the role of patient characteristics and treatment in the prediction of seizure recurrence. This will enable decision-making on the basis of the perceived risk of treatment compared with the benefit of reducing the risk of further seizures in the initial years after diagnosis.


METHODS
A prognostic model was developed based on individual patient data from MESS to enable identification of patients at low, medium, or high risk of seizure recurrence. A split-sample approach was used in which the model was developed on a subsample of the full data and validated on the remainder of the sample. Distinction of the prognostic groups and predictive accuracy of the model were assessed.


FINDINGS
Number of seizures of all types at presentation, presence of a neurological disorder, and an abnormal electroencephalogram (EEG) were significant factors in indicating future seizures. Individuals with two or three seizures, a neurological disorder, or an abnormal EEG were identified as the medium-risk group, those with two of these features or more than three seizures as the high-risk group, and those with a single seizure only as the low-risk group.


INTERPRETATION
The model shows that there is little benefit to immediate treatment in patients at low risk of seizure recurrence, but potentially worthwhile benefits are seen in those at medium and high risk.


Introduction
Epilepsy is a common and diverse disorder with many diff erent causes. Outcomes are varied with 60-70% of newly diagnosed people rapidly entering remission after starting treatment, and 20-30% 1 developing a drugresistant epilepsy with consequent clinical and psychosocial distress. Treatment for most patients is with antiepileptic drugs, which carry risks of acute idiosyncratic reactions, dose-related and chronic toxic eff ects, and teratogenicity. For most patients diagnosed with epilepsy, the benefi ts of treatment will far outweigh the risks associated with treatment. However, for those who have had a single seizure and for those who have seizures with minor symptoms, this risk to benefi t ratio is more fi nely balanced.
A large, multinational, randomised trial (the MRC Multicentre trial for Early Epilepsy and Single Seizures; MESS) was undertaken to assess the benefi ts of starting or delaying treatment after the fi rst reported epileptic seizure. 2 The study showed that although immediate versus delayed treatment with the commonly used antiepileptic drugs carbamazepine and valproic acid reduced the risk of further seizures (hazard ratio for time to fi rst seizure 1·4 [95% CI 1·2-1·7]), there was no evidence of an eff ect on long-term remission rates (absolute diff erence in percentage seizure free between 3 and 5 years -0·2% [95% CI -5·8% to 5·5%]). Decisions on starting treatment should therefore be aff ected by the perceived risk of antiepileptic drug treatment compared with the benefi t of reducing the risk of further seizures in the fi rst 2-3 years after diagnosis. Data from MESS and from other studies [3][4][5][6][7][8] indicate that the risk of further seizures on or off treatment depends on a number of individual characteristics. The decision-making process for embarking on immediate treatment after diagnosis would therefore benefi t from a more formal analysis of the importance of these features.
In this study, we used data for individual patients from the MESS trial to develop a prognostic model for the risk of seizures, and to assess the role of patient characteristics and treatment in this prediction. The model is lent support by a cross-validation sample, enabling assessment and calibration of the model developed.

Patients and procedures
The MESS trial (ISRCTN 98767960) randomised 1443 patients in 13 countries worldwide (50% from the UK). The detailed methods and primary results have been described elsewhere, 2 but a brief summary is provided here. Eligibility for the trial was determined by at least one recent epileptic seizure and by the clinician and patient being uncertain whether to proceed with immediate treatment. Patients were randomly assigned either immediate treatment with an antiepileptic drug or delayed treatment after a time when the clinician and patient agreed treatment was necessary. Whenever antiepileptic drug treatment was given, the choice of drug and treatment regimen was made according to the clinician's usual practice. The MESS trial was approved by the northwest multicentre research ethics committee in the UK and by the ethics committees for participating non-UK centres.
Information about past seizures and neurological and family history, in addition to demographic information and fi ndings of neurological examination, was gathered at baseline. The median age at randomisation was 25·3 years (IQR 17·4-43·4) and the median age reported for fi rst seizure was 24·3 years (16·1-41·9). Electro encephalogram (EEG) and CT scan data were also collected where available. Follow-up information was obtained at 3 months, 6 months, 1 year, and successively at yearly intervals after randomisation, for a median of 4·4 years (3·0-6·3). Details of antiepileptic drug treatment, seizures, and other adverse events were recorded.

Statistical analysis
The complete sample of 1443 individuals was split into a test sample and a validation sample in the ratio 6:4 using a random number generator (Stata 8), stratifi ed by randomised treatment. The test sample was used to develop a prognostic model for the interval from randomisation to fi rst seizure based on Cox regression stratifi ed by treatment allocation. Stepwise regression was used to assess the predictive value of baseline covariates of interest with exclusion at p≥0·1 and inclusion at p≤0·05. The numbers of patients with a neurological disorder, neurological defi cit or impairment, delayed development, or learning disability were small; hence a single variable was used to represent patients with any of these neurological conditions. An abnormal EEG was defi ned as specifi c focal or generalised epileptiform or slow wave abnormality. This defi nition excluded non-specifi c abnormality. Missing values of covariates were imputed using the mean of remaining observations in the sample (mode for categorical variables). Transformation of continuous variables was assessed with Martingale residuals.
Sensitivity analyses were done for only those individuals with complete covariate data, and for only those with an EEG in the interval from 9 months prerandomisation to 3 months post-randomisation (excludes 7% of the test sample with no EEG data and 5% with an EEG outside of this interval). The proportion of individuals who had a further seizure was similar among those with an EEG in and outside of the time period (48%), but was lower in those with missing EEG data (33%). Application of a shrinkage factor 9 to the regression coeffi cients was also considered, to compensate for over-fi tting in the validation sample.
A prognostic index was defi ned as the linear predictor resulting from the fi nal model. The predictive value of this prognostic index is assessed by calculation of a separation statistic, D, which indicates the predictive ability of the index. 10 An optimism-adjusted version of the statistic, D adj , was used to correct for bias when fi tting the model and estimating the separation statistic on the same dataset.
Risk group classifi cations were assigned by use of tertiles of the prognostic index distribution obtained from the model, and probabilities of seizure recurrence by 1, 3, and 5 years then calculated for each of these groups. The prognostic index constructed from the test sample was then applied to the validation sample, enabling assignment of each individual to a specifi c risk group. The predicted risk (as observed in the test sample) of seizure recurrence in each of these groups was compared with the observed seizure recurrence in the validation sample. Kaplan-Meier plots were used to assess the diff erences across risk groups in both the test (predicted risk) and validation (observed risk) samples. The predictive accuracy of the prognostic index in the validation sample was also assessed more formally by use of a censoring adjusted Brier score. 11

Role of the funding source
The funding source had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had fi nal responsibility for the decision to submit for publication.
Number of seizures prerandomisation (log transformation) Years between most recent seizure and randomisation ·· ·· 0·92 (0·84-1·02) p=0·10 First degree relative with epilepsy ·· ·· 1·27 (0·96-1·70) p=0·10 Data are hazard ratios (95% CI), interpretable as the relative change in risk for a one unit increase in the prognostic factor; for log seizures, one unit refers to exp(1)=2·718 seizures. Regression coeffi cients as used in the calculation of the prognostic index can be calculated from hazard ratios as the natural logarithm of the hazard ratio. . Epileptiform abnormality on the EEG (paroxysmal slow-wave abnormality with spiking) added no greater specifi city than the variable "any abnormality", which included also slow-wave disturbance without spikes or sharp waves. Forward stepwise regression was then undertaken to establish the importance of the second set of variables (those signifi cant in univariate analysis at p>0·10), together with interaction terms using the terms identifi ed in model 1 and reconsidered variables excluded from the fi rst set. This resulted in the identifi cation of two further predictive factors: a borderline signifi cant increase in seizure recurrence with a history of epilepsy in a primary relative and a borderline signifi cant decrease with an increasing interval between most recent seizure before randomisation and randomisation itself (model 2). However, this latter association was largely due to two extreme outliers with 13 and 23 years between most recent seizure and randomisation; exclusion of these individuals resulted in this factor not being added to the model. Both models are described in table 2. Sensitivity analysis restricted to individuals with complete covariate information, and to individuals with EEG data from 9 months prerandomisation to 3 months postrandomisation resulted in no change to the models. A prognostic index was calculated, on the basis of model 1, as the sum of the covariate values for a particular patient, weighted by the corresponding estimated regression coeffi cients. So, for an individual with two seizures, an abnormal EEG, and no neurological disorder, the prognostic index (as calculated from table 2) is: (log e 2×log e 1·56)+(1×log e 1·54)+(0×log e 1·35)=(log e 2×0·44)+ (1×0·43)+(0×0·30)=0·74 Since the two variables added at the forward stepwise regression phase were both of borderline signifi cance (model 2), it was decided to omit these from the fi nal prognostic index for simplicity. The separation statistic D adj was 0·77 (95% CI 0·36-0·90). Since the confi dence interval excludes zero, this suggests that the prognostic index has acceptable ability to discriminate between patients' risks of seizure recurrence. The D statistic can also be used to provide guidance on the largest number of prognostic groups that would (with 90% power) maintain signifi cant separation. For our prognostic model, the maximum number of groups likely to maintain reasonable separation is three. 10 Tertiles of the continuous prognostic index described correspond to values of <0·30, <0·50, and ≥0·50-ie, group 1 (low risk) includes individuals with a prognostic index of <0·30, group 2 (medium risk) includes individuals with a prognostic index of 0·30-0·49, and group 3 (high risk) includes individuals with a prognostic index of ≥0·50. Table 3 shows the probabilities of a further seizure by 1, 3, and 5 years in each of these risk groups. No signifi cant diff erence is observed between treatments for low-risk individuals (Log-rank test χ²=1·7, p=0·2), but there is an indication of improvement with immediate antiepileptic-drug treatment for medium-risk and high-risk individuals (7·0, p=0·008; 21·9, p<0·0005, respectively; overall likelihood ratio test for interaction between risk group and treatment 13·27, p=0·001). The risk group for each individual in the validation sample can be calculated by fi rst calculating the prognostic index with their baseline characteristics, then assigning them to a risk group on the basis of the cut-off values identifi ed in the test sample above. Figure 1 shows the observed proportions seizure-free among individuals in each of these predicted risk groups.
Predictive accuracy of the prognostic model in the validation sample was examined by plotting observed proportions of individuals with seizure recurrence within six groups of predicted seizure recurrence (bandwidths 0·2, 0·1, 0·1, 0·1, 0·1, and 0·4) based on the continuous prognostic model 1 in table 2 at 1 and 3 years postrandomisation (fi gure 2). These plots suggest that some shrinkage remains despite adjustment for this, with observed proportions experiencing seizure recurrence less extreme than predicted at very high and very low predicted proportions of seizure recurrence. The censoring adjusted Brier score suggests a degree of success of the model, with a score of 0·23 at 1 year, 0·24 at 3 years, and 0·25 at 5 years (the Brier score ranges from 0 to 1, with a large score indicating poorer predictive accuracy of the model).
Use of the prognostic index in practice requires some simplifi cation; it is useful to rewrite the fi nal model using integer values as shown in table 4. This is derived from the continuous prognostic index obtained from model 1   abnormal EEG or a neurological disorder. Individuals with a prognostic index score of 2-4 have a high risk of recurrence-ie, those with a single seizure, abnormal EEG, and a neurological disorder; those with two to three seizures, an abnormal EEG, or neurological disorder; and any individual with four or more seizures. Once a risk category has been assigned, estimates of the risk of recurrence can be taken from table 3. For example, an individual in the high-risk category has around a 50% chance of seizure recurrence within both 3 and 5 years if antiepileptic drug treatment is started immediately, compared with around a 65% chance of seizures within 3 years (70% within 5 years) if treatment is delayed. By contrast, an individual in the low-risk category has around a 30% chance of seizure recurrence by 3 years (35% by 5 years) under both treatment policies, bearing in mind that there is no signifi cant diff erence between them within the low-risk group.

Discussion
The MESS trial 2 gathered information about a range of clinical factors previously identifi ed as potentially aff ecting seizure recurrence. 12 Our model indicates that risk of seizure recurrence increases with number of seizures at presentation, abnormal EEG, and presence of a neurological disorder. The identifi cation of these factors as the most important in relation to seizure recurrence is lent support by other published work. 13 These fi ndings enabled the development of a relatively simple model, which identifi es patients with low, medium, and high risk of seizure recurrence. The low-risk group comprises only those presenting with a single seizure, no neurological disorder, and a normal EEG. For those in this low-risk group, there is little benefi t to immediate treatment, but potentially worthwhile benefi ts are seen in the medium and high-risk groups.
Internal validation has been investigated using a splitsample approach. Adequate separation of risk groups and reasonable predictive accuracy has been shown, giving some confi dence in the clinical accuracy of the model. External validation on other datasets is needed to provide a more rigorous test of model validity. The model should be used with caution in populations not represented in MESS, and thus should not be relied on in patients with frequent seizures who would generally be advised to start treatment. The model might also be unreliable in patients seen very quickly after a fi rst defi nite seizure.
Our model indicates that individuals with more than one seizure at presentation are at a higher risk of seizure recurrence than those with only one seizure. Among patients with one seizure at randomisation in MESS, the actuarial cumulative percentages having had seizure recurrence at 2 years were 39% and 32% for those randomly assigned delayed and immediate treatment, respectively; for patients with more than one seizure at presentation, these percentages were 61% and 43%, respectively. 2 Trials of antiepileptic-drug treatment versus no treatment after a single tonic-clonic seizure have reported that treatment with carbamazepine or valproic acid reduced the risk of seizure recurrence by 2 years from approximately 60% to 20% (n=91) 7 and that antiepileptic-drug treatment (most commonly phenobarbital) reduced the 2-year risk from 40% to 32% in those randomly assigned immediate treatment (n=419). 8 The risk reduction in MESS is relatively small, and a large proportion of patients with single seizures remained seizure-free at 2 years, even without treatment. This fi nding could be due to diff erences in the inclusion criteria; MESS is unusual in that it included patients with infrequent or minor seizures. Furthermore, noncompliance in the group randomised to no immediate treatment (and the consequent dilution of the observed treatment eff ect) could have varied between trials. Randomisation might also have been relatively delayed in MESS, with 26% randomised within 1 week of the last seizure. By comparison, in the FIRST study, 8 all patients were randomised within a week of the index seizure.
The decision to start antiepileptic-drug treatment after diagnosis of single seizures or epilepsy can be complex because it is aff ected by many social and psychological    factors that are extraneous to the basic risk-benefi t assessment. Avoidance of further seizures might be paramount in someone whose employment is dependent on their ability to drive, whereas a woman of child-bearing age might be unwilling to accept risks to future pregnancies from drug treatment. For these reasons, the main clinical input is the provision of appropriate information to allow informed decision-making by the patient. Indeed, in the absence of this guidance there is unlikely to be good adherence to any policy of immediate drug treatment.
Although patients with frequent seizures at diagnosis are often easy to advise given the risk of further seizures is high, this is not the case for patients with few or infrequent seizures, or seizures with minor symptoms. Here, the risk of future clinically signifi cant seizures is lower and more closely matched by the risks of adverse eff ects associated with antiepileptic-drug treatment. MESS has provided evidence for a signifi cant benefi t in terms of time to fi rst seizure on immediate treatment compared with delayed treatment. 2 Results from MESS also suggest there is no clinically important benefi t from immediate treatment on late outcomes such as terminal remission of 2 years at 3 and 5 years after starting therapy. 2 For this reason, decision-making should be dominated by the degree to which risk of seizures will be reduced by immediate treatment in the fi rst 2 or 3 years after diagnosis.
There is a need to assess the external validity and usefulness of the model in assisting everyday decisionmaking for relevant patients in a new unselected population. The model is, however, consistent with and supportive of current clinical practice. Patients who present with single or infrequent seizures should not be prescribed an antiepileptic drug, but should be referred to a specialist. The predictive model may be a useful adjunct to individual counselling, but will need the support of appropriate expertise to answer any patient questions that might arise from its predictions.