DTI Biomarkers Reach Moderate-Strong Accuracy for Epilepsy Lateralization
Presurgical epilepsy care depends on a difficult question: where is the seizure focus, and is it confined to one side of the brain or spread more broadly? A preprint from the University of Kansas Medical Center suggests diffusion tensor imaging may add useful signal to that decision by pairing white matter microstructure with clinical and demographic data in a machine learning model.
The study examined 334 adults undergoing presurgical epilepsy evaluation and used diffusion tensor imaging measures drawn from 36 bilateral white matter tracts. Fractional anisotropy and mean diffusivity were combined with clinical and demographic variables, then tested across several supervised machine learning approaches. Elastic net logistic regression emerged as the main model for detailed analysis. It classified right-sided, left-sided, and bilateral or generalized epilepsy with AUC values ranging from 0.76 to 0.88.
The best separation came when the model distinguished left-sided epilepsy from bilateral or generalized epilepsy. In that task, the model reached an AUC of 0.88, with accuracy of 0.90, sensitivity of 0.90, and specificity of 0.89. For clinicians weighing surgical candidacy, those are the kinds of numbers that suggest a biomarker may be useful as a complement to standard presurgical workup rather than a replacement for it.
White Matter Microstructure Carries Lateralizing Information
Diffusion tensor imaging captures how water moves through white matter, and those movement patterns reflect the integrity and orientation of axonal pathways. In epilepsy, recurrent seizures and the underlying epileptogenic network can alter connectivity far beyond the cortical region that first generates seizures. A model that reads those changes from multiple tracts may therefore detect asymmetries linked to the side of seizure onset.
The logic is biologically plausible. Focal epilepsy rarely behaves like a single isolated spot; it more often involves a distributed network, with disease-related remodeling in commissural and association fibers that can differ by hemisphere. When those microstructural changes are quantified across many tracts, elastic net regularization can down-weight noisy variables while preserving the most informative patterns. In a presurgical setting, that is exactly the kind of statistical restraint needed when imaging yields many correlated features.
A single sentence captures the most clinically relevant result: the model performed best when it separated unilateral disease from more diffuse epilepsy.
Right-Left Discrimination Was Harder Than Unilateral Versus Diffuse Disease
The weakest task was distinguishing right-sided versus left-sided epilepsy, where the model reached an AUC of 0.76. That drop is not surprising. Lateralization between two unilateral categories is usually more demanding than telling a clearly focal syndrome from a bilateral or generalized one, especially when the underlying imaging signature may overlap across hemispheres or be shaped by seizure burden, medication exposure, and chronic network reorganization.
By contrast, the distinction between unilateral epilepsy and bilateral or generalized epilepsy reached an AUC of 0.83. That pattern suggests the model may be better at detecting whether the brain shows focal versus diffuse lateralized features than at assigning the exact side with high precision in every patient. In practical terms, a tool like this could function as an adjunctive layer of evidence when other preoperative tests are equivocal, particularly in patients whose MRI, EEG, and clinical history do not line up neatly.
The model’s strongest output came from elastic net logistic regression, a method well suited to high-dimensional imaging data because it balances variable selection with shrinkage. In a dataset built from dozens of bilateral tracts, that balance matters. Without it, correlated diffusion measures can overfit quickly and lose reliability in new patients.
Presurgical Evaluation Needs Better Network-Level Tools
Patients with drug-resistant epilepsy often face a long diagnostic path before surgery is considered. Lateralizing the epileptogenic focus determines not only whether surgery is feasible, but also how aggressive resection or ablation can safely be. Imaging biomarkers that help resolve ambiguity may reduce the chance of misclassification when conventional tests point in different directions.
This preprint does not claim that diffusion imaging alone can decide who should go to the operating room. It points to a narrower and more realistic role: adding quantitative white matter information to the presurgical dataset already used by epilepsy teams. Because the model integrated clinical and demographic variables with tract-level diffusion metrics, it reflects how imaging would likely be used in practice rather than as a stand-alone test.
Still, the study has an important boundary. The performance metrics come from retrospective analysis in a single health system, and the authors explicitly call for further studies that incorporate additional presurgical evaluation data into the model. That next step is essential. EEG findings, structural MRI abnormalities, neuropsychological testing, and possibly other imaging modalities could sharpen classification, especially for the harder right-versus-left problem.
What the Preprint Shows and What It Cannot Yet Prove
The dataset included adults evaluated at one academic center, and the work was approved under an institutional review board with retrospective chart review and waiver of consent. The preprint also states that all data used in the analyses are contained in the manuscript tables and supplementary material. One author disclosed support from or paid consulting for Neuropace; the remaining authors reported no conflicts.
Because this is still a preprint, the findings have not yet undergone peer review. That matters in a field where small shifts in preprocessing, feature selection, or cohort composition can change performance materially. Even so, the central signal is coherent: white matter diffusion patterns appear to carry enough lateralizing information to support machine learning classification in presurgical epilepsy evaluation.
A practical takeaway emerges from the numbers. When the question is whether epilepsy is unilateral or more diffuse, the model performed well. When the question is which hemisphere is involved, performance fell. Those are different diagnostic problems, and the discrepancy will shape how any future clinical tool is designed.
The next version of this work will need external validation, broader multicenter data, and direct comparison with the current standard presurgical pathway. Without that, the model remains a promising analytic framework rather than a deployable decision aid.