Identifying epilepsy psychiatric comorbidities with machine learning.

Area of Science:

• Neuroscience
• Psychiatry
• Computational Linguistics

Background:

• Individuals with epilepsy face a higher risk of mental health comorbidities.
• Machine learning (ML) models analyzing spoken language have shown potential in detecting suicidality in adults.
• Comorbid psychiatric conditions are common in adolescents and young adults with epilepsy.

Purpose of the Study:

• To develop ML classifiers using spoken language to identify current or lifetime psychiatric conditions in adolescents and young adults with epilepsy.
• To assess the efficacy of ML models in detecting specific psychiatric disorders and suicidality within this population.

Main Methods:

• Participants over 12 years old with epilepsy were interviewed using the Mini International Neuropsychiatric Interview (MINI) or MINI Kid Tracking.
• Open-ended conversational questions were used to collect spoken language data.
• N-grams and Linguistic Inquiry and Word Count (LIWC) categories informed ML model construction.
• Models were evaluated for four individual disorders and three mutually exclusive groups (no psychiatric disorders, non-suicidal psychiatric disorders, any suicidality) using areas under the receiver operating characteristic curve (AROCs).

Main Results:

• ML classifiers were built from 227 interviews with 122 participants.
• AROCs for models differentiating groups and individual disorders ranged from 57% to 78% (many with P < .02).

Conclusions:

• ML classifiers based on spoken language can reliably identify suicidality and depression in individuals with epilepsy.
• Larger datasets may improve the identification of anxiety and bipolar disorders.
• Spoken language ML analysis presents a viable screening alternative to traditional methods in various healthcare settings.