Deep facial analysis: A new phase I epilepsy evaluation using computer vision.

Area of Science:

• Neurology
• Computer Science
• Artificial Intelligence

Background:

• Semiological observation is crucial for epilepsy presurgical evaluation, but interpreting patient movements presents subjective challenges.
• Existing computer-based epilepsy monitoring methods largely ignore facial movements, an area with limited research advancements.

Purpose of the Study:

• To develop and evaluate deep learning approaches for automatic extraction and classification of semiological patterns from facial expressions in epilepsy patients.
• To address limitations in current analytical methods by focusing on the understudied area of facial movements during epilepsy monitoring.

Main Methods:

• Proposed two deep learning models: a landmark-based and a region-based approach, to quantitatively analyze facial semiology in mesial temporal lobe epilepsy (MTLE) patients.
• Utilized a dataset from the Mater Advanced Epilepsy Unit, employing multifold cross-validation and leave-one-subject-out cross-validation for evaluation.

Main Results:

• The landmark-based approach showed promise for frontal facial views.
• The region-based approach achieved higher accuracy (95.19% test accuracy, 0.98 AUC) with spatiotemporal features, especially in extreme head positions.
• Leave-one-subject-out validation for the region-based approach yielded lower accuracy (50.85%) due to data limitations.

Conclusions:

• Deep learning models demonstrate potential for quantifying ictal facial movements in MTLE patients.
• These automated methods can enhance presurgical epilepsy evaluation by standardizing analysis, mitigating bias, and improving clinical decision-making.
• Computer-aided diagnosis may aid in preventing erroneous localization and guiding surgical interventions.