- Top of page
Purpose: Although several independent predictors of seizure freedom after temporal lobe epilepsy surgery have been identified, their combined predictive value is largely unknown. Using a large database of operated patients, we assessed the combined predictive value of previously reported predictors included in a single multivariable model.
Methods: The database comprised a cohort of 484 patientswho underwent temporal lobe surgery for drug-resistant epilepsy. Good outcome was defined as Engel class 1, one year after surgery. Previously reported independent predictors were tested in this cohort. To be included in our final prediction model, predictors had to show a multivariable p-value of <0.20.
Results: The final multivariable model included predictors obtained from the patient's history (absence of tonic–clonic seizures, absence of status epilepticus), magnetic resonance imaging [MRI; ipsilateral mesial temporal sclerosis (MTS), space occupying lesion], video electroencephalography (EEG; absence of ictal dystonic posturing, concordance between MRI and ictal EEG), and fluorodeoxyglucose positron emission tomography (FDG-PET; unilateral temporal abnormalities), that were related to seizure freedom in our data. The model showed an expected receiver-operating characteristic curve (ROC) area of 0.63 [95% confidence interval (CI) 0.57–0.68] for new patient populations. Intracranial monitoring and surgery-related parameters (including histology) were not important predictors of seizure freedom. Among patients with a high probability of seizure freedom, 85% were seizure-free one year after surgery; however, among patients with a high risk of not becoming seizure-free, still 40% were seizure-free one year after surgery.
Conclusion: We could only moderately predict seizure freedom after temporal lobe epilepsy surgery. It is particularly difficult to predict who will not become seizure-free after surgery.
Epilepsy surgery is an effective treatment for medically intractable epilepsy, especially in patients with temporal lobe epilepsy (TLE). After TLE surgery, 60% to 70% of patients become seizure-free, and 90% of patients achieve a worthwhile reduction in seizure severity (Engel et al., 1998, 2003). The presurgical work-up for epilepsy surgery is stepwise and complex, and contradictory findings from standard tests [history, seizure semiology, electroencephalography (EEG), and magnetic resonance imaging (MRI)] with regard to lateralization or localization of the seizure focus necessitate additional tests of increasing invasiveness and cost [e.g., ictal single photon emission computed tomography (SPECT), positron emission tomography (PET), intracranial EEG recordings]. To be able to inform candidates for TLE surgery about their chances of postoperative seizure freedom, it is important to define which characteristics are true predictors of seizure freedom after surgery. This requires a multivariable study approach (Harrell et al., 1996). The ultimate goal would be to develop a simple clinical prediction model or rule to predict the chance of seizure freedom after surgery for individual patients undergoing TLE surgery.
Previous studies of predictors of postoperative seizure freedom using multivariable analysis differ in their methodology and results (Armon et al., 1996; Berg et al., 1998; Radhakrishnan et al., 1998; Jeong et al., 1999, 2005; Hennessy et al., 2001a, 2001b; Clusmann et al., 2002, 2004; Tonini et al., 2004; Janszky et al., 2005, 2006; Spencer et al., 2005; Cohen-Gadol et al., 2006; Jeha et al., 2006; Kelemen et al., 2006; Yun et al., 2006; Malmgren et al., 2007). Although potential independent predictors have been identified, the predictive value of combinations of these independent predictors (i.e., the value of these predictors combined in a single prediction model) has been investigated in only one study, which included patients with all types of epilepsy and not only TLE (Armon et al., 1996). The aim of the present study was therefore to use a large homogeneous database of patients who underwent TLE surgery to quantify the predictive accuracy of the combination of previously reported predictors of seizure freedom.
- Top of page
Of the 484 patients, 356 patients [incidence 74%, 95% confidence interval (CI) 0.69–0.77] were seizure-free (Engel class I) one year after surgery.
Univariable associations between predictors and outcome are presented in Table 2. Based on restricted cubic splines, two continuous predictors, “age at time of surgery” and “duration of epilepsy,” were not linearly related to the outcome but had to be modeled as square root terms. The other continuous predictors were linearly related to the outcome and could be modeled as such.
Table 2. Univariable associations of each potential predictor with Engel class 1 (yes/no) as outcome
|Outcome: Engel class 1||N 356 (74%)||Odds ratio||95% CI||p-value|
|Female sex||248 (51%)||0.87||0.58–1.30||0.48|
|Febrile seizures||180 (37%)||0.90||0.59–1.36||0.61|
|Epilepsy duration (in years)||21.6 ± 11.4||0.98||0.97–1.00||0.05|
|Age at start epilepsy (in years)||11.8 ± 9.3 ||0.99||0.98–1.02||0.87|
|No tonic–clonic seizures||190 (39%)||2.39||1.52–3.76||<0.001 |
|No status epilepticus||427 (88%)||1.75||0.98–3.12||0.06|
|Total IQ (score)||33.3 ± 11.3||0.99||0.99–1.01||0.74|
|Age at surgery (in years)||103.0 ± 18.4 ||0.98||0.96–1.00||0.03|
|Left-sided surgery||232 (48%)||0.89||0.60–1.34||0.59|
|Abnormal MRI||443 (92%)||1.14||0.60–2.18||0.68|
|Space occupying lesion on MRIa||102 (21%)||1.50||0.88–2.55||0.13|
|Concordance MRI and EEG||255 (53%)||1.37||0.92–2.06||0.13|
|No ictal limb dystonia||221 (46%)||1.21||0.81–1.82||0.36|
|No extratemporal semiology||408 (84%)||0.91||0.52–1.61||0.76|
|No bilateral interictal spikes||368 (76%)||0.96||0.60–1.55||0.87|
|No extratemp interictal spikes||439 (91%)||1.01||0.51–2.03||0.97|
|No start ictal EEG posterior temporal||462 (96%)||1.05||0.40–2.73||0.93|
|Ipsilateral delayed rhythmical θ onset||132 (27%)||0.77||0.49–1.19||0.24|
|Concordance interictal and ictal EEG|| 93 (19%)||0.85||0.52–1.41||0.53|
|FDG-PET unilateral temporal||136 (28%)||1.42||0.71–2.81||0.32|
|No intracranial monitoring||414 (86%)||0.99||0.52–1.93||0.47|
|Resection size (in cm)||4.3 ± 1.3||1.01||0.86–1.17||0.96|
|Postoperative discharges||249 (51%)||1.18||0.79–1.77||0.43|
|MTS on histology||269 (56%)||0.81||0.54–1.22||0.31|
|No cortical dysgenesis on histology||444 (92%)||1.21||0.60–2.46||0.60|
The overall model included all predictors from history, MRI, and video EEG monitoring. Ipsilateral delayed rhythmical θ onset in ictal EEG was excluded from the model based on the sign OK method. Backwards stepwise exclusion further reduced the starting model to six predictors of seizure freedom (model 1, Table 3): age at time of surgery, absence of tonic–clonic seizures or status epilepticus in the patient's history, presence of ipsilateral MTS or a space occupying lesion on the MRI, and absence of ictal dystonic posturing. None of the extra predictors proposed by the members of the Dutch Collaborative Epilepsy Surgery Program were of added predictive value to this reduced model.
Table 3. Models including the important predictors of postoperative seizure freedom
| ||Model 1||Model 2||Model 3||Model 4|
|Odds ratio (95% CI)||p-value||Odds ratio (95% CI)||p-value||Odds ratio (95% CI)||p-value||Odds ratio (95% CI)||p-value|
|No tonic–clonic seizures||2.24 (1.40–3.58)|| 0.001||2.31 (1.44–3.70)|| 0.001||2.32 (1.45–3.73)||<0.01 ||2.32 (1.45–3.73)||<0.01 |
|No status epilepticus||1.62 (0.88–2.98)||0.12||1.54 (0.83–2.84)||0.17||1.52 (0.82–2.82)||0.18||1.45 (0.78–2.71)||0.24|
|Age at surgerya||0.84 (0.68–1.04)||0.10||0.83 (0.67–1.03)||0.09||0.83 (0.67–1.03)||0.09||0.84 (0.68–1.05)||0.12|
|MTSb on MRI||1.63 (1.03–2.58)||0.04||1.61 (1.02–2.57)||0.04||1.59 (0.99–2.55)||0.06||1.80 (1.05–3.08)||0.03|
|Space occupying lesion on MRI||1.67 (0.92–3.02)||0.09||1.68 (0.93–3.03)||0.09||1.63 (0.89–2.99)||0.11||1.57 (0.85–2.90)||0.15|
|No ictal dystonic posturing||1.34 (0.87–2.05)||0.19||1.36 (0.89–2.10)||0.16||1.35 (0.88–2.08)||0.18||1.36 (0.87–2.12)||0.17|
|FDG-PET unilateral temporal||nac|| ||1.47 (0.95–2.29)||0.09||1.47 (0.95–2.29)||0.09||1.50 (0.96–2.35)||0.07|
|Intracranial monitoring performed||na|| ||na|| ||1.14 (0.62–2.07)||0.68||1.14 (0.62–2.13)||0.67|
|Resection size||na|| ||na|| ||na|| ||1.02 (0.86–1.20)||0.86|
|Postoperative discharges||na|| ||na|| ||na|| ||1.15 (0.76–1.77)||0.51|
|MTSb on histology||na|| ||na|| ||na|| ||0.78 (0.46–1.34)||0.37|
|No cortical dysgenesis on histology||na|| ||na|| ||na|| ||1.17 (0.55–2.49)||0.68|
FDG-PET abnormalities were an additional predictor of seizure freedom [odds ratio (OR) = 1.47; 95% CI 0.95–2.29; p-value: 0.09] (model 2), whereas intracranial monitoring (OR = 1.14; 95% CI 0.62–2.07; p-value = 0.68) (model 3), and operative predictors (model 4) did not add any predictive information.
The Hosmer-Lemeshow test indicated good calibration, with a p-value of 0.79 for model 1, 0.35 for model 2, 0.47 for model 3, and 0.57 for model 4. This was confirmed by the calibration plots (not shown).
Model 2, based on predictors from the patient's history, and MRI, video EEG, and FDG-PET findings yielded the best prediction model, with an ROC area of 0.66 (95% CI 0.60–0.70). After correction for optimism, based on bootstrapping, this ROC area was reduced to an ROC area of 0.63 (95% CI 0.57–0.68), a value that can be expected if this model is used with other similar patient populations.
Table 4 shows the number of patients with and without seizure freedom after one year across the probability categories predicted by model 2. This table shows how many patients with a certain probability of becoming seizure-free actually did or did not become seizure-free one year after surgery. The observed incidence of seizure freedom (whether patients actually became seizure-free) increased from 40% in the group with the lowest probability according to our model to 85% in the highest probability group. The risk of not becoming seizure-free ranged from 15% in the group with the highest probability of seizure freedom to 60% in the lowest. This means that 40% of patients with the highest risk of not achieving seizure freedom were nevertheless seizure-free one year after surgery.
Table 4. Number (%) of patients with or without seizure freedom after one year over the probability categories estimated by model 2a
|Estimated probability based on model 2 in Table 3||Seizure freedom N = 356||No seizure freedom N = 128|
|<0.45 (N = 5; 1% of 484)|| 2 (40%)|| 3 (60%)|
|0.45–0.60 (N = 52; 11%)||31 (59%)||21 (41%)|
|0.60–0.70 (N = 112; 23%)||74 (66%)||38 (33%)|
|0.70–0.80 (N = 161; 33%)||118 (73%) ||43 (26%)|
|>0.80 (N = 154; 32%)||131 (85%) ||23 (15%)|
- Top of page
We assessed all 22 predictors found in earlier multivariable studies on seizure freedom after TLE surgery (Armon et al., 1996; Berg et al., 1998; Radhakrishnan et al., 1998; Jeong et al., 1999, 2005; Hennessy et al., 2001a, 2001b; Clusmann et al., 2002, 2004; Tonini et al., 2004; Janszky et al., 2005, 2006; Spencer et al., 2005; Cohen-Gadol et al., 2006; Jeha et al., 2006; Kelemen et al., 2006; Yun et al., 2006; Malmgren et al., 2007) and identified seven predictors of postoperative seizure freedom; namely younger age at time of surgery, a history without tonic–clonic seizures, a history without status epilepticus, MRI with ipsilateral MTS, MRI with space occupying lesion, no dystonic posturing during the seizure, and unilateral temporal abnormalities on FDG-PET. The other predictors from the basic diagnostic work-up, additional diagnostic tests, and operative data did not contribute to the prediction of postoperative seizure freedom. Our final model included all predictors reported by Jeong et al. (2005), Spencer et al. (2005), and Janszky et al. (2006).
We reanalyzed the data on the hypothesis that patients with MTS or a space-occupying lesion might have a more similar predictive value, while patients with other abnormalities have a different (smaller) risk of becoming seizure-free. To this end, we took “MTS and/or space-occupying lesion” as one predictor and “other abnormalities on MRI, other than MTS and space-occupying lesion” as a separate predictor. This analysis did not yield different results: the combined predictor “MTS and/or space-occupying lesion” would be included in the final model, but this model yielded a lower ROC area than in our original analysis.
As model 4 shows, resection size was no additional predictor. As this predictor is only meaningful in patients who underwent a tailored resection, we performed a subgroup analysis in the 400 patients who underwent a tailored resection, which gave the same result.
Our study presents an overall predictive value, that is a measure of how the use of such a model would discriminate between postoperative seizure freedom or not. This overall predictive value of the combination of predictors was moderate, with an ROC area 0.63. This means that we were unable to formulate a simple and stable prediction rule to predict seizure freedom that could be used to inform patients. The model can be used to indicate risk categories for postoperative seizure freedom. However, it performs insufficiently to be used for individual patients to discriminate between becoming and not becoming seizure-free.
Of earlier studies, only the one by Armon et al. (1996) included a measure of the performance of their model in predicting postoperative seizure freedom. Armon et al. (1996) found a Somers' D coefficient of 0.47 or a ROC area of 0.74 without correction for optimism, on the basis of five preoperative predictors: ipsilateral imaging abnormality, ipsilateral EEG localization (ictal and interictal), intracranial EEG recordings, temporal lobe resection, and age (Armon et al., 1996). Since their study involved patients who had undergone temporal or extratemporal resections, their model is not directly comparable to ours. However, the predictors “ipsilateral imaging abnormality” and “age” were also included in our model.
Unfortunately, other studies predicting postoperative seizure freedom did not present the overall accuracy of their model (nor could this be reconstructed with the data provided). The wide variation of preoperative predictors reported in the literature and the moderate overall predictive value of our own model indicate that it is difficult to predict postoperative seizure freedom one year after TLE surgery. In prognostic medical research, as in all areas of life, prediction becomes more difficult the further ahead the outcomes we want to predict (Moons et al., 2002). This means that the presence or absence of particular predictors in an individual patient cannot directly be associated with an increased or decreased chance of becoming seizure-free after surgery.
To appreciate our results, some methodological aspects need to be discussed. First, the study outcome measure was Engel class I, one year after surgery. We reanalyzed the data with the outcome absolute seizure freedom (Engel class IA) one year after surgery, which led to the same results (i.e., the same independent predictors were identified). Secondly, we wanted to include ancillary tests, such as FDG-PET, which were not performed in all patients. FDG-PET was performed in 188 of 484 patients, mostly when MRI and video EEG monitoring results were inconclusive. Imputation of FDG-PET results in patients in whom FDG-PET was not performed, as described earlier, enabled us to assess the independent value of FDG-PET in the complete patient population (Uijl et al., 2007). We reached the same conclusion when we restricted our analysis to the subgroup of 188 patients in whom FDG-PET was performed. Thirdly, the predictors were necessarily reduced to essentials for categorization. Since the number of predictors that can be included in a prognostic model is limited, we only included previously reported predictors of seizure freedom. This obviously does not fully reflect the subtle nuances of interpretation that often arise in clinical practice, and the model cannot comprise all possible information; these complexities necessarily have been obscured.
In conclusion, whereas the results of many preoperative tests in TLE surgery have a statistically significant association with postoperative seizure freedom, in combination they are only moderate predictors of postoperative seizure freedom. It is particularly difficult to predict the absence of postoperative seizure freedom. Therefore, currently available data do not yet allow the development of a robust prediction rule for postoperative seizure freedom. More refined (software) analysis of existing tests, new diagnostic tests such as EEG-functional MRI (EEG-fMRI), and even genetic analysis, may provide future opportunities to improve the prediction of postoperative seizure freedom.