Beyond Mean Pooling: Statistical Meta-Aggregation for Subject-Level EEG Classification in Parkinson's Disease
- 1University of Calgary
Abstract
Electroencephalography (EEG) provides a non-invasive, temporally precise tool for studying neural alterations associated with Parkinson's disease (PD). When labels are defined at the subject level but inputs are epoch level, a key design choice is how epoch-wise predictions are aggregated to a single subject-level decision. Most prior work uses mean pooling, which assumes all epochs contribute equally and independently. This study proposes a statistical meta-aggregation strategy that represents each subject by 14 descriptors of the epoch-wise prediction distribution, including measures of central tendency, dispersion, shape, and high-confidence concentration. Using a self-supervised contrastive encoder trained on task-evoked EEG from the Simon conflict task, and a sample of 90 subjects including 57 PD and 33 control, this work evaluates the proposed framework under subject-wise five-fold cross-validation. Replacing mean pooling with statistical meta-aggregation lifts subject-level ROC-AUC from 0.655 to 0.872 for logistic regression and from 0.724 to 0.840 for support vector machines, without modifying the encoder or the classifier family. Averaging predictions across three independent self-supervised pretraining seeds yields competitive performance and improved balanced accuracy. These results suggest that, in this dataset, the way epoch-level predictions are summarized at the subject level has a substantial effect on final PD classification performance.
Related
- Window leakage in clinical EEG machine learning: matched subject-independent evaluation across ADHD and Parkinson cohorts
- MangoSwin-attention: a swin transformer with attention pooling for generalizable mango leaf disease classification
- The Parkinson's disease-associated A53E mutation decreases alpha-synuclein aggregation and phosphorylation in vivo
- Interpretable YOLO and mask-guided pooling for maize leaf disease classification in complex field backgrounds
- Application of Bayesian-Derived Gene Interaction Network and Subnetwork for Classification of Idiopathic Parkinson Disease with Biomarkers
BibTeX
@misc{langrodi2026beyond,
title = {{Beyond Mean Pooling: Statistical Meta-Aggregation for Subject-Level EEG Classification in Parkinson's Disease}},
author = {P. Taghipour Langrodi and A. Khodadadi and F. Rezaei and Z. Anvarian and R. C. Sotero and S. Madadi},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.13.759017},
url = {https://www.biorxiv.org/content/10.64898/2026.10.13.759017v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Beyond Mean Pooling: Statistical Meta-Aggregation for Subject-Level EEG Classification in Parkinson's Disease
AU - P. Taghipour Langrodi
AU - A. Khodadadi
AU - F. Rezaei
AU - Z. Anvarian
AU - R. C. Sotero
AU - S. Madadi
PY - 2026
DA - 2026/10/13
PB - bioRxiv
DO - 10.64898/2026.10.13.759017
UR - https://www.biorxiv.org/content/10.64898/2026.10.13.759017v1
AB - Electroencephalography (EEG) provides a non-invasive, temporally precise tool for studying neural alterations associated with Parkinson's disease (PD). When labels are defined at the subject level but inputs are epoch level, a key design choice is how epoch-wise predictions are aggregated to a single subject-level decision. Most prior work uses mean pooling, which assumes all epochs contribute equally and independently. This study proposes a statistical meta-aggregation strategy that represents each subject by 14 descriptors of the epoch-wise prediction distribution, including measures of central tendency, dispersion, shape, and high-confidence concentration. Using a self-supervised contrastive encoder trained on task-evoked EEG from the Simon conflict task, and a sample of 90 subjects including 57 PD and 33 control, this work evaluates the proposed framework under subject-wise five-fold cross-validation. Replacing mean pooling with statistical meta-aggregation lifts subject-level ROC-AUC from 0.655 to 0.872 for logistic regression and from 0.724 to 0.840 for support vector machines, without modifying the encoder or the classifier family. Averaging predictions across three independent self-supervised pretraining seeds yields competitive performance and improved balanced accuracy. These results suggest that, in this dataset, the way epoch-level predictions are summarized at the subject level has a substantial effect on final PD classification performance.
ER -