Abstracts

Rationale: Anti-seizure medications (ASMs) are successful in controlling seizures for approximately 60% of patients newly treated for epilepsy. However, with more than 30 approved ASMs worldwide, there is no reliable way to predict which would be most effective for a given patient. Consequently, it may take several years to find an effective ASM regimen for an individual with epilepsy. We aimed to develop an artificial intelligence (AI) model to predict seizure freedom in response to ASMs and to potentially suggest the most suitable medication at the onset of treatment.


Methods: Consecutive adults ( >18 years) with newly diagnosed epilepsy who had been treated with one or two ASMs as monotherapy from 2015 to 2024 were retrospectively recruited from the patient database of the Alfred Hospital (Melbourne, Australia) (Table 1). A second regimen was administered (either as a monotherapy or as an adjunctive medication), if the first ASM failed to yield an effective outcome (at least 12 months of seizure freedom after starting medication). Patients who underwent 3T MRI (Siemens, Skyra) were eligible for inclusion. The fusion AI model consisted of an 18-layer 3D-ResNet (inputs: three baseline MR images), a long short-term memory (LSTM) recurrent neural network (inputs: the two most recent regimens), and a dual linear neural network (inputs: clinical characteristics and EEG findings) to predict the probability of seizure outcome. The ASMs with less than 3 occurrences within the sample population were omitted and the corresponding patients were excluded. Patients without any high quality/resolution MRIs (T1, T2, and FLAIR images) were also excluded. The probability of seizure freedom per ASM was predicted by the fusion model and thresholded to be compared with the ASM seizure outcome. To train the model, 50 iterations of 5-fold cross-validation was performed. The accuracy, F1-score, and AUC (the area under the curve) for the ROC (receiver operating characteristic) curve were calculated and averaged among the 5 folds.


Results: One-hundred and thirteen patients were included in the analysis, of which 68 patients were seizure free and 45 were not. The fusion model accuracy achieved: 84% and 75%, F1-score: 84% and 74%, and AUC: 89% and 73%, on training and test data, respectively. Figure 1 shows the ROC curves for each of 5 folds in training and validating steps with their AUC. It suggests that the proposed fusion model can predict the ASM leading to seizure freedom for roughly as many as 75% of newly diagnosed epilepsy.


Conclusions: Response to initial ASMs may be predicted using a fusion AI model that integrates ASM information with patient’s characteristics and MRI images. This approach holds promise for personalized treatment strategies, leading to improved treatment outcomes.


Funding: None