Predictive model for epileptogenic tubers from all tubers in patients with tuberous sclerosis complex based on ¹⁸F-FDG PET: an 8-year single-centre study

This study was approved by our institutional review board, and all subjects signed informed consent forms. We selected consecutive TSC patients who presented to the comprehensive epilepsy center from June 2013 to July 2021. The inclusion criteria were (i) patients diagnosed with TSC according to the diagnostic criteria in 2021 [3]; (ii) intractable epilepsy and seizures occurring at least twice per month on average during the 6 months prior to surgery; (iii) comprehensive preoperative evaluations and epilepsy surgery at our comprehensive epilepsy centre; (iv) ¹⁸F-FDG PET examination completed more than 48 h since the last seizure; and (v) at least 1 year of follow-up. The exclusion criteria were (i) patients with other specific neurological abnormalities, mainly including encephalitis, hydrocephalus, intracerebral haemorrhage or cerebrovascular disease; (ii) contraindications for imaging examinations; and (iii) who dropped out of the study or loss to follow-up. The study flowchart is presented in Fig. 1.

Clinical characteristics (including data on sex, age at onset, frequency of seizures, number of antiseizure medications [ASMs], and number of tubers) were collected. Comprehensive preoperative examinations, including neurological physical examinations, genetic tests, MRI, ¹⁸F-FDG PET/CT, scalp video-EEG and stereoelectroencephalography (SEEG), were performed by neurosurgeons, neurologists and neuroradiologists to determine whether and how to perform epilepsy surgery on the TSC patients. We first chose a noninvasive method to identify the epileptogenic tubers based on our and other previous studies [4, 18, 19]. An epileptogenic tuber was defined as (i) a cortical tuber with focal ictal and interictal epileptiform discharges in the same brain region on scalp EEG with consistent seizure symptoms; (ii) an outstanding cortical tuber with focal ictal or interictal epileptiform discharges in the same brain region on scalp EEG with consistent seizure symptoms; or (iii) an outstanding cortical tuber with focal ictal and interictal epileptiform discharges in the same brain region on scalp EEG. Outstanding cortical tubers described were defined as having a size > 3–4 cm and containing a nidus of calcification according to imaging results. For those patients with multiple epileptogenic cortical tubers or discrepancies among clinical symptoms, imaging results and electrophysiology findings, intracranial electrodes were invasively implanted to record electrical activity and identify epileptogenic tubers. In addition, resected “non-epileptogenic tubers” in this study are adjacent to epileptogenic tubers, exposed to the same surgical field with the epileptogenic tubers and located in the non-functional areas, and these tubers are large, prominent or calcified tubers but without epileptic discharges. These tubers have the potential to become epileptogenic tubers, strongly suggest pharmacoresistant epilepsy in TSC, influence postoperative seizure freedom and should be considered good indicators for epilepsy surgery [4, 20]. Patients finished their last follow-up in July 2022. The clinical characteristics of the patients are shown in Table 1.

The criteria are summarized as follows: at least 4 h of fasting before FDG PET/CT, blood glucose level at the time of FDG injection < 11.1 mmol/L, and PET acquisition was done 1 h after the injection of 3 MBq/kg of FDG. In addition, MRI scans were performed with a 3.0-T system with a vendor-produced 32-channel head coil. Patients were in an awake, resting state, and no patients had seizures during the imaging scans. Image registration was utilized to spatially align the ¹⁸F-FDG PET and T2Flair MRI images using CT images as a reference, and the registered image showed the spatial correspondence between PET and MRI images (Additional file 1: Fig. S1, Step 1–2). Then, regions of interests (ROIs) were delineated on the T2Flair MRI images and the delineated ROIs also applied to the PET images (Additional file 1: Fig. S1, Step 3). The quantitative indices of the delineated ROIs were further extracted based on the PET images (Additional file 1: Fig. S1, Step 4). In each delineated ROI, 19 quantitative indices that reflect the clinical significance of the PET image were identified [17, 21‐24]. A detailed explanation of these indices is shown in Additional file 1: Tab. S1. Researchers blinded to the clinical and imaging features delineated the ROIs and analysed the imaging data independently. All processing procedures were performed under the guidance of experienced PET/CT specialists.

Logistic regression (LR), linear discriminant analysis (LDA) and artificial neural network (ANN) models were constructed using R 4.2.1 software (R Foundation for Statistical Computing, Vienna, Austria). ROC analysis was used to confirm the identification performance of all models. The relationship between the PET quantitative indices and the probability of identifying an epileptogenic tuber was visualized at the bottom of the nomogram. Calibration curves were used to assess the agreement between the predicted probability and the actual proportion, and the Hosmer–Lemeshow test was used to estimate the calibration of the nomogram model. In addition, decision curve analysis (DCA) and clinical impact curve (CIC) were performed to evaluate the utility of the nomogram.

The proteins of the resected brain tissues from TSC patients were extracted, and their concentration was measured using a BCA protein assay. The extracted protein was mixed with 5 × loading buffer, electrophoresed in 10% sodium dodecyl sulfate polyacrylamide gel and transferred onto polyvinylidene difluoride membranes. The membranes were then blocked in 5% skim milk for 2 h at room temperature. Then, the membranes were incubated with the primary antibodies overnight at 4 °C, including anti-mTOR (1:1000, Abcam) and anti-GAPDH (1:1000, Abcam). After washing, the membranes were incubated with anti-rabbit or anti-mouse secondary antibodies for 1 h at room temperature. Specific protein bands were visualized in the membranes using the chemiluminescence method, and the optical density of the protein bands was analysed in Image-Pro Plus software.

Continuous variables were compared by the unpaired two-tailed T tests. The relationship between quantitative indices and epileptogenesis was evaluated using logistic regression analysis. Bivariate correlation analyses were carried out using Spearman’s rank correlation test. Receiver operating characteristic curve (ROC) analysis was performed to evaluate the performance of the models. The area under the curve (AUC) was calculated to illustrate the diagnostic power of the models. A nomogram was built to help guide clinicians, and DCA was performed to evaluate the utility of the nomogram. R 4.2.1 software was used for data analysis. Significance was set at P < 0.05.

Data from 61 TSC patients who underwent epilepsy surgery in our centre were collected. Eighteen patients were excluded based on the exclusion criteria. Forty-three patients (20 female, 23 male) were ultimately enrolled in this study (Fig. 1).

The clinical characteristics of the TSC patients were collected (Table 1). According to gene testing, 8 patients had a TSC1 variant, 29 had a TSC2 variant and 6 patients had no gene variants. The average age at onset was 20.16 months (range 1–156 months). Seizures occurred every day in 26 patients, per week in 10 patients and at least twice per month in 7 patients. The average number of ASMs was 2.19 (range 1–4). Eight patients had 1–3, twenty-six patients had 4–6 and nine patients had more than 6 cortical tubers. According to the ictal scalp EEG findings, 10, 12, 14 and 7 patients had focal, lateralized, multifocal and generalized epileptic discharges, respectively. Furthermore, 16 patients underwent subdural EEG monitoring, and 13 patients underwent SEEG monitoring. Tuberectomy or multiple tuberectomy was performed on 17 patients, lobectomy or multiple lobectomy was performed on 11 patients, and lobectomy plus tuberectomy was performed on 15 patients. One-year follow-up was performed for all patients, and 32 patients were seizure-free. Three-year follow-up was performed for 38 patients, and 26 patients were seizure-free. An over-5-year follow-up was performed for 32 patients, and 21 patients were seizure-free.

Overall, 235 tubers (71 epileptogenic and 164 non-epileptogenic) from 43 patients were identified, and their PET quantitative indices were extracted and analysed. The results showed that the SUVmean, SUVmax, volume, total lesion glycolysis (TLG), third quartile, upper adjacent and standardized added metabolic activity (SAM) were associated with epileptogenic tubers (Table 2), and the SUVmean, SUVmax, volume, TLG and root-mean-square deviations (RMSD) values were different between epileptogenic tubers and non-epileptogenic tubers (Additional file 1: Tab. S2). Based on these results, we further selected the SUVmean, SUVmax, volume and TLG indices to investigate the association of PET quantitative indices and epileptic characteristics in TSC patients with a single epileptogenic tuber and found that SUVmean, SUVmax and TLG were negatively associated with seizure frequency and duration (Additional file 1: Fig. S2). Abnormal activation of the mTOR pathway is the pathogenic mechanism of epilepsy in TSC [25]. Intriguingly, SUVmean and SUVmax were negatively associated with activated mTOR expression in resected epileptogenic tubers (Additional file 1: Fig. S3). The above results indicated that the PET quantitative indices have potential predictive value for epileptogenic tubers.

The PET quantitative indices from 235 tubers were used as the training set. The AUCs of the training set in the LR, LDA and ANN predictive models were 0.7706 (95% CI, 0.70–0.83), 0.7506 (95% CI, 0.68–0.82) and 0.7425 (95% CI, 0.67–0.82), respectively (Fig. 2), indicating that the LR model performed better in predicting epileptogenic tubers than the LDA and ANN models (Fig. 2). The LR model was internally evaluated based on ten-fold cross-validation on discrimination and calibration, and the results showed that it was well calibrated (Hosmer‒Lemeshow goodness-of-fit p value = 0.7). In addition, the multicollinearity of the PET quantitative indices was evaluated by the variance inflation factor (VIF), and the VIFs were less than 5, indicating that there was no multicollinearity in these models. Therefore, we selected the LR model to evaluate the values of the quantitative indices in predicting epileptogenic tubers.

Furthermore, a nomogram was constructed incorporating epileptogenic tuber-associated PET quantitative indices to guide clinicians in identifying epileptogenic tubers in TSC patients (Fig. 3A). Each patient was given points for each quantitative indices, and the higher the total points, the more likely the tuber was to be epileptogenic. In addition, calibration curves demonstrated that the nomogram had a similar performance to the ideal model and confirmed the good predictive capacity of the nomogram (Fig. 3B). DCA (Fig. 4A) and CIC (Fig. 4B) were established to confirm the clinical applicability of the nomogram. DCA showed that the prediction of epileptogenic tubers with the nomogram provide greater net benefits when the high-risk threshold probability was > 30%. CIC was used to predict the stratification of the probability of epileptogenic tubers for a sample size of 1000 and suggested that the number of identified epileptogenic tubers was close to the actual number of epileptogenic tubers when the threshold probability was higher than 0.4, and the cost-to-benefit ratio was 0.75.

New PET data of 32 tubers from 7 patients were used as the testing set to evaluate the predictability of the LR model. The AUC of the LR model was 0.8502 (95% CI, 0.71–0.99), and the accuracy was 78.13% (Table 3). Furthermore, we evaluated the prognostic value of the LR model in the long-term outcomes of cortical tubers based on the 1-, 3- and 5-year follow-up data. In the 1-year follow-up of 74 epileptogenic tubers, the AUC value was 0.7805 (95% CI, 0.71–0.85), and the accuracy was 79.15%. In the 3-year follow-up of 77 epileptogenic tubers, the AUC value was 0.8066 (95% CI, 0.74–0.87), and the accuracy was 80.43%. In the 5-year follow-up of 80 epileptogenic tubers, the AUC value reached 0.8172 (95% CI, 0.75–0.87), and the accuracy was 79.15% (Table 3). These findings showed the capability of the LR model to predict the long-term outcomes of cortical tubers.