Montage-agnostic federated learning for privacy-preserving Alzheimer’s disease classification from heterogeneous multi-site EEG
Year: 2026
Type: article
Abstract: Background. Federated learning (FL) trains shared models across institutions without moving raw data, which is attractivefor clinical electroencephalography (EEG) where data are scarce and privacy-sensitive. A practical obstacle has receivedlittle attention: clinical sites record EEG with different hardware and electrode montages, so the input feature space itselfdiffers across clients and standard aggregation cannot be applied. A second, quieter problem affects the evidence basefor EEG machine learning. Because a single recording yields thousands of short epochs, splitting epochs (rather thansubjects) between training and test sets leaks subject identity and inflates reported accuracy. This study asks a singlequestion: once heterogeneous montages are harmonized and the evaluation is made leakage-free, what do clinical sitesactually gain from federated EEG learning? Methods. We present MontageFL, a montage-agnostic harmonization pipelinethat maps any electrode configuration to a fixed 88-dimensional feature vector through region-based aggregation, combiningspectral biomarkers (Stream A, 32-d) with covariance-derived features (Stream B, 56-d). The harmonized representation letsa 19-channel site and an 11-channel site train together. We benchmark centralized, local, FedAvg, and FedProx training onAlzheimer’s disease (AD) versus healthy control (HC) classification using the Miltiadous et al. (2023) dataset (65 subjects),under a strictly subject-level evaluation (StratifiedGroupKFold, no subject in both train and test) across four seeds, withleave-one-site-out (LOSO) cross-montage testing. We additionally quantify the effect of epoch-level leakage by re-evaluatingthe identical pipeline under both splitting schemes. Results. Under leakage-free subject-level evaluation, all training paradigmsperformed comparably: on the 19-channel site FedProx reached 67.1±14.3% accuracy (AUC 0.731) and local training 68.7%;on the 11-channel site FedProx reached 73.9% and local training 77.0%. No method significantly outperformed another(FedProx vs. local p=0.14, Cohen’s d=−0.27; FedProx vs. FedAvg p=0.33). Federated and local models were statisticallyindistinguishable, and federation introduced no accuracy penalty. The same pipeline evaluated with epoch-level splitting reported89–97% accuracy, 15 to 30 percentage points higher, illustrating the magnitude of leakage-driven inflation. Conclusions.Montage-agnostic harmonization makes heterogeneous-hardware EEG sites interoperable for privacy-preserving collaboration,and federation matches centralized and local training without sharing raw data. On datasets of this size, however, federationyields no accuracy advantage over local training, not even for a deliberately data-poor site. There, naive size-weightedaggregation significantly degraded performance (−9.4 percentage points, p<10−6); equal client weighting fully repaired thisaggregation artifact (+9.1 points, p<10−8) but only restored parity with local training, not an advantage. Realistic accuracy isalso lower than many headline figures reported in the EEG-FL literature. We suggest that subject-level evaluation should bereported explicitly in this field, and we release a leakage-free benchmark to support that practice.
Source: Scientific Reports
Authors: Bahaa Masry, Anas Shahin, George Karraz
Institutions: Damascus University, Syrian Virtual University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7202074445