Abstract
Electroencephalography (EEG) is a low-cost, non-invasive candidate biomarker for Parkinson’s disease (PD), and deep-learning models report high balanced accuracy on a benchmark that pools four public PD datasets. We show that, in this benchmark, pooled accuracy is strongly influenced by site-associated label imbalance: a diagnostic site-prior null that uses no EEG signal (predicting each dataset’s majority class) reaches 0.90 segment-level balanced accuracy, comparable to published models. Evaluating instead with three-site leave-one-dataset-out (LODO) over the both-class PD datasets, fixed-threshold balanced accuracy is substantially degraded. Our analysis suggests that this degradation includes a major threshold/calibration component rather than a complete loss of discriminative information: supervised scores rank PD versus healthy controls on unseen sites at ROC-AUC 0.76 ± 0.03, and a fully deployable decision threshold selected on the training sites reaches 0.64 ± 0.03 balanced accuracy. Finally, under the architecture, data, and scale tested, self-supervised pretraining provides no observed cross-site gain: frozen linear probes on encoders pretrained on the four datasets or on a large disjoint clinical corpus (TUH) reach cross-site AUC of 0.58 and 0.53 respectively, and fine-tuning matches but does not improve over training from scratch, with no data-efficiency advantage. We will release the evaluation tooling and recommend reporting cross-site PD-EEG with LODO, threshold-independent metrics, and an explicit site-prior null.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_069.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Edward_Lue_Chee_Lip1
BibTex
@InProceedings{LipEdw_ABenchmark_MICCAISAT2026,
author = { Lip, Edward Lue Chee AND Mai, Van AND Neema, Saanvi AND Jameson, Alexander AND Torbus, Karolina AND Suresh, Jithin},
title = { { A Benchmark Audit of Site Confounds, Calibration, and Self-Supervision in Cross-Dataset Parkinson’s EEG Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}