Surgery for Epilepsy
Using Artificial Intelligence To Improve Surgical Outcomes

January 4, 2025

Surgery stops seizures for 70% of temporal lobe epilepsy patients who have it. But can we detect who will benefit and who wont, to avoid unneccessary surgery?


Around 39 million people worldwide suffer from epilepsy. This disease has profound physical, psychological and social consequences. For around 30% of epilepsy patients, medication is unable to control the seizures and for these patients, the most common treatment is brain surgery to remove the damaged tissue. This procedure is effective and leads to an improved quality of life, and a meaningful reduction in seizures for approximately 70% of patients. However, there can be serious side effects, such as reductions in memory, language deficits and motor complications. Although patients are thoroughly examined prior to surgery to assess their suitability for the procedure, for around 30% of patients do not obtain reductions in seizure frequency.

The decision to undergo brain surgery is a difficult decision for patients suffering from epilepsy. We want to help patients and their doctors have more certainty around this decision, and improve outcomes for people suffering with epilepsy.

We are using PET and MRI brain imaging from patients to try to learn what patterns in the images predict good surgical outcome, and use this as a tool for helping neurologists and patients decide whether to proceed to surgery. 

We recently discovered that the presence of abnormalities observed in PET images on the opposite side of the brain from the tissue to be resected was associated with poor surgical outcomes, but only for patients whose epliepsy orginates on the right side of the brain! 

We suspect that there are other patterns in the images presently unknown which could predict whether the surgery will be successful. We are attempting to find these patterns using machine learning.

At present we have collected and screened data of 145 epilepsy patients who had surgery and underwent PET scans prior to surgery, with about 250 more yet to be screened. This is the largest known data set of this kind.

So far data from a subset of 82 patients has been analysed. Summary measures from the PET and MRI images have been extracted and used to predict surgical outcome with an efficiency of 0.75(±0.09) using support vector machines. This analysis was limited to patients who had MRI scans before and after surgery, which substantially reduced our sample size. An additional limitation was that the full images were not input into the machine learning algorithm, only summary measures of the images.

We intend to use AI to discover features on MRI and PET scans predictive of good surgical outcomes (i.e. reduction in seizure frequency). We will use unsupervised generative deep learning methods to encode PET and MRI images to and from a latent space. We will predict surgical outcomes from the latent encodings, and use the generative model to answer – “How would this patient’s MRI or PET have to change in order for them to have a good surgical outcome?” By comparing synthetic images to the original images, we can identify features predictive of good surgical outcomes. We refer to this method as the creation of Generative Visual Rationales. 

Without AI, we can only make tests with a priori hypothesis and prior domain knowledge. With AI we automatically extract image features without having to define these manually, potentially identifying features that we have not thought of or identified before. 


©2020 by Ben Sinclair.