Development and validation of a predictive model for the diagnosis of neural antibody-mediated epilepsy/ seizure in patients with new-onset seizure or established epilepsy.

Purpose: Currently, the diagnosis of neural antibody-mediated epilepsy/seizure (NAME/S)relies heavily on neural antibody testing, which is time-consuming, costly and introduces diagnostic delays. A statistical tool to predict the probability of a patient with NAME/S is lacking. We aimed to construct a predictive model to help clinicians expedite the diagnostic process.

Methods: We retrospectively recruited subjects (206 in the development group and 62 in the validation group) with new-onset seizures or established epilepsy suspected to have presented with antibody-mediated seizures between January 2014 and December 2019. We collected data about demographics, medical history, clinical manifestations and follow up. Binary logistic regression was used to select potential predictors for the construction of a predictive model. Five-fold cross and bootstrap validation were applied to avoid overfitting. Concordance index, calibration plots and decision curve analysis were used to assess its performance.

Results: The model, incorporating presence/absence of tumour, psychiatric/cognitive/emotional changes, language disturbances, sensory auras, tonic-clonic seizures, multiple seizure events, hyponatremia and MRI inflammation, was visualized as a nomogram. The crude and adjusted concordance indices were both 0.88 with a cut-off value of 0.62, sensitivity of 83.2 % and specificity of 77.4 %. The slope and intercept of the calibration curve were 0 and 1, respectively. The model also showed good performance in the validation group with a concordance index of 0.82, cut-off value of 0.33, sensitivity of 75.5 % and specificity of 73.1 %. The slope was 0.86 and the intercept was 0.039. Decision curve analysis showed that the model was useful with an optimal threshold probability of ＞4 % in both groups.

Conclusions: Despite limitations such as sample volume and selection bias in subject enrolment, this model may be used to estimate the individualized probability of having NAME/S, deserving further exploration and validation.