Resting EEG Complexity Predicts but Does Not Generalize for iTBS Response
According to a preprint posted on Research Square and indexed in PubMed, Matthew Ning et al. evaluate whether baseline resting-state EEG complexity and transcranial magnetic stimulation measures can predict neurophysiological responses to a single session of intermittent theta-burst stimulation (iTBS) applied to primary motor cortex (Res Sq preprint, 2025 Sep 24; PubMed PMID 41041549). The authors used statistical reliability analyses and supervised machine-learning models that integrated baseline rsEEG features and TMS-evoked measures including motor-evoked potentials (MEPs) and TMS-evoked potentials (TEPs). The preprint reports that while some predictive signal was present within-sample, the models failed to generalize to independent validation, as reflected in the manuscript title and reported results in the preprint. The work is a preprint and has not been peer reviewed (PubMed).
What happened
According to the Research Square preprint indexed on PubMed (Res Sq preprint, posted 2025 Sep 24; PubMed PMID 41041549), Matthew Ning et al. analyzed whether baseline resting-state EEG (rsEEG) complexity and TMS-evoked measures predict neurophysiological responses to a single session of intermittent theta-burst stimulation (iTBS) over primary motor cortex. The authors report applying statistical and reliability analyses and training supervised machine-learning models that combined rsEEG features with TMS outcome measures including motor-evoked potentials (MEPs) and TMS-evoked potentials (TEPs). The preprint reports predictive performance within the original sample but states that models "fail to generalize," a finding highlighted in the manuscript title (Res Sq preprint; PubMed). The manuscript is a preprint and has not been peer reviewed (PubMed).
Technical details
Editorial analysis - technical context: The study's feature set, as reported, centers on measures of EEG complexity from resting recordings plus evoked-response metrics from single-pulse TMS, combined in supervised learning pipelines. The preprint emphasizes testing beyond single-visit, within-sample correlations by attempting independent validation, a step often missing in prior rTMS ML work according to the authors' framing in the abstract (Res Sq preprint; PubMed). The paper does not, in the PubMed abstract, provide full performance metrics or external cohort details in the indexed summary; readers should consult the full preprint for model types, cross-validation strategy, sample sizes, and reported effect sizes (Res Sq preprint).
Context and significance
Industry context: Machine-learning research on neuromodulation biomarkers routinely encounters limited sample sizes, large inter-individual variability, and overfitting risks. The preprint's central negative finding about generalization aligns with broader reproducibility concerns in physiological biomarker ML. For the research community, that pattern highlights the difference between within-sample signal discovery and deployable, generalizable predictive models.
What to watch
For practitioners and researchers: monitor whether the authors release full code and data, whether they report clear external validation cohorts or pre-registered replication, and whether peer review alters reported model performance. Independent replication on larger, multi-site datasets will be the key indicator of whether rsEEG complexity can serve as a robust biomarker for iTBS responsiveness.
Scoring Rationale
The preprint tackles an important methodological question for neuromodulation biomarkers using ML, which is notable for researchers in applied neurotech. Impact is limited by preprint status, domain specificity, and the reported failure to generalize, which reduces immediate practical utility.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
