Predicting the cause of seizures using features extracted from interactions with a virtual agent

Area of Science:

• Neurology
• Artificial Intelligence
• Computational Linguistics

Background:

• Transient Loss of Consciousness (TLOC) is frequently misdiagnosed, leading to delays in appropriate specialist care.
• Current clinical decision tools often struggle to differentiate between the primary causes of TLOC: epilepsy, functional/dissociative seizures (FDS), and syncope.
• Previous research indicates distinct linguistic patterns in spoken accounts of epileptic versus FDS seizures.

Purpose of the Study:

• To explore the feasibility of enhancing TLOC diagnosis by integrating automated analysis of patient-reported symptoms and spoken TLOC descriptions.
• To develop a clinical decision tool that improves upon existing methods for differentiating common TLOC causes, particularly epilepsy and FDS.
• To assess the predictive accuracy of combining a symptom questionnaire with linguistic features derived from spoken TLOC narratives.

Main Methods:

• Participants completed an online questionnaire (iPEP) and interacted with a virtual agent (VA) to provide spoken descriptions of their TLOC events.
• Support Vector Machines (SVM) were employed for classification, utilizing features from the iPEP and three sets of linguistic features: formulation effort, semantic word categories, and parts of speech (verb, adverb, adjective usage).
• A nested leave-one-out cross-validation strategy was used to train and evaluate the SVM models.

Main Results:

• The iPEP questionnaire alone achieved a diagnostic accuracy of 65.8%.
• Incorporating linguistic features derived from the VA interaction significantly improved diagnostic accuracy to 85.5%.
• The enhanced model demonstrated a marked improvement in differentiating between epilepsy and FDS, addressing a key limitation of previous tools.

Conclusions:

• Automated analysis of spoken TLOC descriptions, when integrated with symptom data via an online application and VA, can substantially improve diagnostic accuracy.
• This approach offers a promising method for enhancing clinical decision support tools for TLOC.
• The findings suggest potential for improved clinical stratification, guiding appropriate referrals for cardiological or neurological investigations and management.