02 Predicting the cause of TLOC using an automated analysis of interactions with a virtual agent

Objectives/Aims A clinical decision tool for Transient Loss of Consciousness (TLOC) could reduce misdiagnosis rates and waiting times. Most clinical decision tools fail to stratify between the three most common causes of TLOC (epilepsy, functional (dissociative) seizures, and syncope) or are hindered by the challenging differentiation between epilepsy and FDS. Based on previous research describing differences in spoken accounts of epileptic and nonepileptic seizures, this study explored the feasibility of predicting the cause of TLOC by combining the automated analysis of patient-reported symptoms and spoken TLOC descriptions. Method Participants completed an online web application that consisted of a 34-item medical history and symptom questionnaire (iPEP) and interaction with a virtual agent (VA) that asked eight questions about the most recent experience of TLOC. Support Vector Machines (SVM) were trained using different combinations of features and nested leave-one-out cross validation. The iPEP provided a baseline performance. Three language-based feature sets were created: features designed to measure formulation effort, features that measured the proportion of words from different semantic categories, and features based on verb, adverb, and adjective usage. Two methods of integrating the iPEP and language features were compared. Method one involved training a single SVM model using all features and all diagnoses. Method two used a ‘model stacking’ approach whereby predictions of epilepsy or FDS from the iPEP model were passed into a second stage language analysis to improve this differential diagnosis. Results 76 participants completed the application (Epilepsy = 24, FDS = 36, syncope = 16). Only 61 participants also completed the VA interaction (Epilepsy = 20, FDS = 29, syncope = 12). The iPEP model accurately predicted 65.8% of diagnoses. For the binary classification between epilepsy and FDS, the three language feature sets predicted the diagnosis with an accuracy between 75.5–85.7%. Combining the iPEP and language features resulted in an overall accuracy of 59% for the first integration method and 85.5% for method two (model stacking). Conclusion These findings suggest that an automated analysis of TLOC descriptions collected using an online web application and VA could improve the accuracy of current clinical decisions tools for TLOC and facilitate clinical stratification processes (such as the appropriate referral to cardiological versus neurological investigation and management pathways). Future research should aim to improve the baseline performance of the iPEP and explore methods for analysing TLOC descriptions from patients with syncope that can improve the identification of this diagnostic group.