Tumor-masked structural connectomics for assessing tract-specific white matter disruption associated with glioma aggressiveness: an interpretable machine learning pilot study.
Authors
Affiliations (10)
Affiliations (10)
- Faculty of Engineering, Universidad Andrés Bello, Santiago, Chile.
- Biomedical Imaging Center, School of Medicine, Pontificia Universidad Católica de Chile, Santiago, Chile.
- Millennium Institute for Intelligent Healthcare Engineering, Santiago, Chile.
- Institute of Physics, Pontificia Universidad Católica de Chile, Santiago, Chile.
- Department of Psychiatry, School of Medicine, Pontificia Universidad Católica de Chile, Santiago, Chile.
- Radiology Department, Hospital Carlos Van Buren, Valparaíso, Chile.
- Neurosurgery Department, Hospital Carlos Van Buren, Valparaíso, Chile.
- Biomedical Engineering School, Faculty of Engineering, Universidad de Valparaíso, Valparaíso, Chile.
- Millennium Institute for Intelligent Healthcare Engineering, Santiago, Chile. [email protected].
- Biomedical Engineering School, Faculty of Engineering, Universidad de Valparaíso, Valparaíso, Chile. [email protected].
Abstract
To assess the feasibility of personalized tumor-masked structural connectomics combined with interpretable machine learning (ML) for differentiating low-grade gliomas (LGG) from high-grade gliomas (HGG), and to identify candidate tract-specific connectomic biomarkers of glioma aggressiveness. Thirty-five patients with histopathologically confirmed gliomas (20 LGG, 15 HGG) underwent preoperative diffusion tensor imaging and T1-weighted MRI. Tumor-masked probabilistic tractography was used to reconstruct individualized structural connectomes based on 50 white matter (WM) tract labels. We evaluated 296 graph-theoretical features using sequential feature selection and Random Forest classification within a strictly isolated nested cross-validation framework. We used SHAP analysis for model interpretation. Sensitivity analyses compared clinical-radiological, connectomic, and combined feature sets. The leak-free pure connectomics model achieved a mean ROC-AUC of 0.814 ± 0.327 (95% CI, 0.500-1.000), with a macro-F1 score of 0.786 ± 0.265. Restricting the analysis to non-zero connectomic features yielded a mean AUC of 0.833 ± 0.167. The clinical-radiological model achieved a mean AUC of 0.817 ± 0.149, while the combined clinical-radiological and connectomic model showed the highest point estimate (AUC, 0.933 ± 0.109; 95% CI, 0.798-1.068). SHAP analysis highlighted the sagittal stratum, medial lemniscus, and superior fronto-occipital fasciculus among the most informative WM pathways. Tumor-masked structural connectomics combined with interpretable ML demonstrated preliminary feasibility for identifying candidate imaging biomarkers associated with glioma aggressiveness in neuroradiological practice. These findings are exploratory and represent internal estimates from a single cohort; external validation in independent, multi-center cohorts is required before any potential clinical application.