A deep representation learning model to predict response to vagus nerve stimulation

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Pediatric epilepsy affects many children, with vagus nerve stimulation (VNS) being a common but inconsistently effective treatment.
• Current methods cannot predict VNS outcomes, leading to surgeries without benefit for approximately half of pediatric patients.
• Preoperative T1-weighted magnetic resonance imaging (T1w) shows potential for predicting VNS response but faces challenges due to high dimensionality.

Purpose of the Study:

• To develop a predictive model for VNS treatment outcomes in pediatric epilepsy using preoperative T1w MRI data.
• To overcome the limitations of traditional predictive modeling with high-dimensional imaging data.
• To improve clinical decision-making for VNS implantation in children with epilepsy.

Main Methods:

• A deep representation learning model, VQ-VNS, was developed to predict VNS response using preoperative T1w images (n=263).
• The model was pre-trained on a large dataset of 7433 T1w images to learn compact anatomical representations.
• Performance was evaluated on the largest pediatric VNS cohort (n=1046), comparing VQ-VNS predictions to clinical data-based predictions.

Main Results:

• Presurgical clinical data showed poor predictive power for VNS response (AUC 0.54, p>0.99).
• The VQ-VNS model accurately predicted VNS response with an AUC of 0.73 (p=0.007).
• Model predictions highlighted disruptions in serotonin-rich brain regions and large-scale network connectivity in non-responders.

Conclusions:

• VQ-VNS, a deep learning model, effectively predicts VNS outcomes in pediatric epilepsy using routine structural MRI.
• The model offers biological interpretability, identifying network disruptions associated with non-response.
• This approach enhances clinical decision-making, potentially improving treatment efficacy and reducing healthcare costs.