Clinical efficacy and predictive model development for levetiracetam in children with newly diagnosed epilepsy of unknown etiology

Background: Levetiracetam (LEV) is one of the most widely used antiseizure medications (ASMs) in pediatric epilepsy owing to its broad-spectrum antiseizure efficacy, favorable safety profile, and good tolerability. However, considerable interindividual variability exists in treatment response to LEV. This study aimed to evaluate the efficacy and safety of LEV monotherapy in children with newly diagnosed epilepsy of unknown etiology, to identify clinical predictors associated with LEV treatment response, and to develop and validate a machine learning (ML)-assisted clinical risk assessment scale based on routinely available clinical information, thereby providing an objective and quantitative tool to support individualized treatment decision-making in pediatric epilepsy. Methods: This single-center cohort study enrolled 115 children with newly diagnosed epilepsy of unknown etiology who were treated at The First Medical Center of Chinese PLA General Hospital between October 2021 and October 2024. All patients received LEV monotherapy and were followed up for 12 months. Treatment success was defined as seizure freedom for at least 12 consecutive months from the initiation of LEV. Clinical characteristics were collected, relevant predictors were selected, ML algorithms were applied to construct models, and SHapley Additive exPlanations (SHAP) were used to interpret feature importance. A clinical risk assessment scale was ultimately developed. Results: Among the 115 children, the 12-month seizure-freedom rate was 73.9% (85/115). Adverse drug reactions (ADRs) occurred in 11.3% (13/115) of patients, all of which were Grade 1–2 in severity with no serious adverse events. Statistical analysis identified six clinical features for inclusion in the risk scale: the time from first seizure to treatment initiation, seizure frequency, seizure type (Tonic-Clonic), age at onset, history of febrile seizures, and sex. The scale demonstrated an area under the receiver operating characteristic curve (AUC) of 0.812 [95% confidence interval (CI): 0.747–0.877]. Conclusions: LEV monotherapy demonstrated good efficacy and tolerability in children with newly diagnosed epilepsy of unknown etiology. The time from first seizure to treatment initiation was the strongest independent clinical predictor of LEV efficacy, and its positive predictive effect may be attributable to the high proportion of self-limited epilepsy with centrotemporal spikes (SeLECTS) in this cohort. The clinical risk assessment scale developed in this study requires only six routine clinical parameters obtainable at the visit, demonstrates good discriminative validity and clinical utility, and may serve as a decision-support tool for individualized LEV treatment in tertiary epilepsy centers and specialist settings where complete etiological workup can be performed at the time of initial clinical encounter. Its applicability in resource-limited settings where etiological classification cannot be promptly established warrants further investigation. External validation in multicenter, large-sample prospective cohorts is warranted.