With keen interest we read Ooi et al.'s recent publication,1 and we congratulate the authors on their impressive work. The authors show that machine-learning models incorporating advanced quantitative MRI features (e.g., cortical folding and interhemispheric asymmetry) meaningfully improve seizure recurrence predictions after a first seizure compared with clinical features alone (AUC 0.65 vs. 0.57), in a carefully selected cohort of adult patients with nondiagnostic MRI and EEG. Importantly, by excluding patients with epileptogenic MRI findings (and epileptiform EEG abnormalities), the authors avoid incorporation bias—the methodological error of using predictor variables that directly define the outcome, which can artificially inflate model performance.2, 3 This makes the study particularly compelling, as it offers insight into novel imaging-based predictors beyond traditional clinical and demographic variables, which are known to have limited predictive value.4, 5 Because MRI and EEG findings are an integral part of clinical decision-making in practice5-8 previously published models have included them as predictors.9, 10 One of these models,10 as mentioned by Ooi et al.,1 included patients with epileptogenic MRI lesions known to be associated with the outcome (seizure recurrence), and the high discriminatory ability of that model may be attributed to this.1 Similarly, our model published in 2018 on childhood epilepsy prediction showed that including EEG findings—including epileptiform EEG abnormalities—as a predictor substantially increased model performance compared to clinical features alone (AUC 0.86 vs. 0.67).9 Since the presence of epileptiform EEG or epileptogenic MRI findings after a first seizure is often considered sufficient for an epilepsy diagnosis, including them as predictors is methodologically suboptimal.2, 5, 11 We therefore agree with Ooi et al.'s approach of excluding patients with these diagnostic features when assessing the added predictive value of potential new diagnostic or prognostic biomarkers. In this letter we seek to point out these contrasting modeling strategies—specifically, the choice to include or exclude diagnostic features—and to encourage researchers and clinicians to explicitly balance the risk of incorporation bias against clinical applicability. Excluding diagnostically decisive features during model development inevitably limits a model's ability to capture the full range of relevant information. However, it also produces more realistic predictions, enabling discovery of new biomarkers.12 We advocate for broader dialogue to harmonize study designs in first-seizure research, and echo Ooi et al.'s view that integrating novel biomarkers13 may meaningfully advance clinical prediction models. We thank the authors for their important contribution to the field. None of the authors has any conflict of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
Kampman et al. (Mon,) studied this question.