This study’s main strength is its deliberately conservative, data-centric evaluation of HFO-derived features in a large, open iEEG cohort, with a clear metadata-only baseline and repeated stratified CV logistic regression to quantify incremental predictive value. The conclusion is appropriately framed: pooled prediction is weak (AUC ~0.56 baseline; ~0.63 with refined HFO features), and the most informative result is the pathology-dependent behavior—showing larger (still moderate) gains in FCD and a heterogeneous “Other” group. The authors also avoid common overclaims by explicitly arguing against universal-biomarker narratives and by noting that coefficient plots are descriptive rather than confirmatory. The main uncertainty is whether the observed subgroup “lifts” reflect true pathology-conditional signal versus instability from smaller effective sample sizes, residual confounding, and dataset/derivative heterogeneity (different recording methods, sampling rates, detector differences, SOZ/resection definition variability, and missingness). AUCs in subgroups remain near-chance (e.g., 0.54 in FCD), so clinical utility is not established; additionally, it is unclear whether calibration, decision-curve utility, confidence intervals, and rigorous nested CV (including any feature selection/tuning steps) were used to prevent optimistic estimates. Overall, the paper’s modest conclusion—heterogeneity matters and pooled models dilute effects—is justified by the reported effect sizes, but stronger evidence would require tighter control of leakage/confounding and more complete uncertainty quantification and reproducibility details (exact preprocessing, inclusion/exclusion, and code/config to regenerate the derivatives-to-features pipeline).