EUVIMEDEUVIMED entdecken
Crossref

Privacy-Preserving Federated Learning with Explainable AI for Multi-Site Autism Spectrum Disorder Biomarker Discovery

Anusha Ravankol, Nandula Anuradha

Pertanika Journal of Science and Technology · 30.10.2026 · 34 (5)

DOI: 10.47836/pjst.34.5.17 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Pertanika Journal of Science and Technology, online erschienen am 30.10.2026. Verfasst von Anusha Ravankol und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study develops a privacy-preserving federated learning framework, coupled with explainable AI, for collaborative discovery of autism spectrum disorder (ASD) biomarkers without centralising sensitive patient data. A federated system spanning three simulated hospital sites that fuses structural and functional MRI via cross-modal attention is evaluated with differential privacy applied to both model aggregation (Gaussian mechanism, calibrated with a tight Rényi differential privacy accountant) and biomarker discovery (Laplace mechanism, Fed-SHAP). On a held-out test set drawn from a 173-subject cohort (and never used for training or hyperparameter selection), the proposed model achieves 92.6% accuracy (95% bootstrap CI: [81.5%, 100.0%]; permutation test p = 0.001 against a chance-level mean of 50.4%) at a composed privacy guarantee of ε = 1.0, δ = 10-5 over 10 communication rounds. Leave-one-site-out validation, in which each site is held out entirely as an external test set, yields a mean accuracy of 86.7% (SD = 4.4%), providing a more conservative estimate of cross-institution generalisation. Fed-SHAP, a privacy-preserving feature-importance approximation rather than a direct computation of Shapley values, ranked 20 candidate functional-connectivity biomarkers, which we mapped to AAL-116 anatomical regions and networks; the most frequently implicated networks were cerebellar and association-cortex regions (25.0% each), followed by subcortical/limbic (15.0%) and visual (12.5%) regions. Because individual rankings showed only 20.6% bootstrap stability, these candidate biomarkers are exploratory and are reported at the network level rather than as fixed individual connections.

Zitieren

APA

Ravankol, A., & Anuradha, N. (2026). Privacy-Preserving Federated Learning with Explainable AI for Multi-Site Autism Spectrum Disorder Biomarker Discovery. Pertanika Journal of Science and Technology, 34(5). https://doi.org/10.47836/pjst.34.5.17

Vancouver

Ravankol A, Anuradha N. Privacy-Preserving Federated Learning with Explainable AI for Multi-Site Autism Spectrum Disorder Biomarker Discovery. Pertanika Journal of Science and Technology. 2026;34(5). doi:10.47836/pjst.34.5.17

BibTeX
@article{Ravankol2026,
  author = {Ravankol, Anusha and Anuradha, Nandula},
  title = {Privacy-Preserving Federated Learning with Explainable AI for Multi-Site Autism Spectrum Disorder Biomarker Discovery},
  journal = {Pertanika Journal of Science and Technology},
  year = {2026},
  volume = {34},
  number = {5},
  doi = {10.47836/pjst.34.5.17}
}
Kostenlos starten →