Background Transient loss of consciousness (TLOC) is a common primary care presentation. 90% are due to syncope (S), epilepsy (E), or psychogenic non-epileptic seizures (PNES). Misdiagnosis and delayed diagnosis is common. We explore symptoms and witness observations that can classify patients with likely diagnoses of E, S, or PNES.

Methods Patients with objectively-documented diagnoses of E, S, or PNES, and an attack witness, were invited to complete a questionnaire (capturing medical history, 86 peri-episodal experiences, and 31 witness observations). Iterative feature selection identified questions strongly predictive of diagnosis; a random forest trained on these classified patients into likely diagnoses of E, S, or PNES.

Results 249 patients (86 E, 79 s, 84 PNES) were randomly assigned to training or validation in a 2:1 ratio. Feature selection identified 36 highly-predictive questionnaire items. The classifier correctly diagnosed 86% of patients in validation. 100% of S were correctly diagnosed, 85.7% E and 75% PNES. A simpler 12-feature model correctly classified 76.7% of cases (E: 75%; S: 92.3%; PNES: 65.6%).

Conclusions TLOC-associated symptoms and manifestations can contribute to a decision rule for primary/emergency care, assisting triage and referral. Determining a diagnostic pre-test probability from TLOC features can aid interpretation of investigation abnormalities of uncertain significance.