Improving diagnostic procedures for epilepsy through automated recording and analysis of patients’ history

Transient loss of consciousness (TLOC) is a time-limited state of profound cognitive impairment characterised by amnesia, abnormal motor control, loss of responsiveness, a short duration and complete recovery. Most instances of TLOC are caused by one of three health conditions: epilepsy, functional (dissociative) seizures (FDS), or syncope. There is often a delay before the correct diagnosis is made and 10-20% of individuals initially receive an incorrect diagnosis. Clinical decision tools based on the endorsement of TLOC symptom lists have been limited to distinguishing between two causes of TLOC. The Initial Paroxysmal Event Profile (iPEP) has shown promise but was demonstrated to have greater accuracy in distinguishing between syncope and epilepsy or FDS than between epilepsy and FDS. The objective of this thesis was to investigate whether interactional, linguistic, and communicative differences in how people with epilepsy and people with FDS describe their experiences of TLOC can improve the predictive performance of the iPEP. An online web application was designed that collected information about TLOC symptoms and medical history from patients and witnesses using a binary questionnaire and verbal interaction with a virtual agent. We explored potential methods of automatically detecting these communicative differences, whether the differences were present during an interaction with a VA, to what extent these automatically detectable communicative differences improve the performance of the iPEP, and the acceptability of the application from the perspective of patients and witnesses. The two feature sets that were applied to previous doctor-patient interactions, features designed to measure formulation effort or detect semantic differences between the two groups, were able to predict the diagnosis with an accuracy of 71% and 81%, respectively. Individuals with epilepsy or FDS provided descriptions of TLOC to the VA that were qualitatively like those observed in previous research. Both feature sets were effective predictors of the diagnosis when applied to the web application recordings (85.7% and 85.7%). Overall, the accuracy of machine learning models trained for the threeway classification between epilepsy, FDS, and syncope using the iPEP responses from patients that were collected through the web application was worse than the performance observed in previous research (65.8% vs 78.3%), but the performance was increased by the inclusion of features extracted from the spoken descriptions on TLOC (85.5%). Finally, most participants who provided feedback reported that the online application was acceptable. These findings suggest that it is feasible to differentiate between people with epilepsy and people with FDS using an automated analysis of spoken seizure descriptions. Furthermore, incorporating these features into a clinical decision tool for TLOC can improve the predictive performance by improving the differential diagnosis between these two health conditions. Future research should use the feedback to improve the design of the application and increase perceived acceptability of the approach.