Letter to the Editor on “Individualized structural network deviations predict surgical outcome in mesial temporal lobe epilepsy: a multicenter validation study”

Dear Editor, Feng et al report a multicenter MRI-based machine learning model that uses individualized structural covariance networks (iSCN) to predict surgical outcome in mesial temporal lobe epilepsy (mTLE)[1]. In a field where 30–40% of patients remain non–seizure-free after anterior temporal lobectomy despite optimal candidate selection[2,3], the shift from a purely focal “epileptogenic zone” concept toward network-level biomarkers is both timely and clinically relevant. This commentary is prepared in accordance with the TITAN 2025 guidance to ensure transparent and responsible reporting of AI-related methods[4]. The key methodological advance is the derivation of a patient-specific iSCN from routine preoperative T1-weighted MRI using a normative template, then partitioning it into a “surgery network” and a “surgically spared network” based on a standardized resection mask. Feng et al. show that non–seizure-free (NSF) patients exhibit significantly greater iSCN abnormalities in the spared network, particularly within contralateral frontal, temporal, insular and limbic modules, whereas the surgery network does not differ between outcome groups[1]. This supports the notion that persistent epileptogenic networks outside the resection zone critically determine surgical failure. Using support vector machines, contralateral spared iSCN features predicted 1-year seizure outcome with an accuracy of 82% and AUC 0.81 in the primary cohort, with external validation accuracies of 80 and 88% (AUC 0.80–0.82) in two independent centers[1]. Default mode and frontoparietal networks contributed most to classification, echoing prior work linking network hub abnormalities and temporal lobe epilepsy outcome[5–7]. Importantly, the authors package their model into an open-source “Epilepsy Surgical Prediction Toolbox,” outlining a realistic workflow from preoperative MRI to virtual resection and prognostic readout, which is a major translational strength. Several caveats warrant caution before routine adoption. The study is retrospective and restricted to three Chinese centers; generalizability to other populations, scanners and surgical practices remains uncertain. One external cohort is small (n = 17), so performance estimates there are imprecise. A group-level resection mask was used rather than patient-specific postoperative masks, which may blur the distinction between resected and spared tissue. Moreover, the model is not compared head-to-head with simpler predictors (clinical variables, standard MRI, or established connectomic scores), nor is there an impact analysis showing that model-guided decisions improve outcomes or change resection strategy. Nonetheless, this work meaningfully advances the concept that contralateral and spared network pathology – rather than local temporal lobe damage alone – drives postoperative seizure persistence. For epilepsy surgeons and multidisciplinary teams, iSCN-based tools are best viewed as adjuncts to, not replacements for, comprehensive presurgical evaluation. The next step should be prospective, ideally multicenter, studies embedding iSCN-guided risk stratification into standard work-up to determine whether individualized structural network biomarkers can safely refine candidate selection and tailoring of resection in mTLE.