Style transfer mapping with ensemble learning for cross-subject EEG-based neuropsychiatric disease recognition.
4 affiliations
- 1Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin, China.
- 2Institute of Brain Science and Brain-inspired Research, Shandong First Medical University & Shandong Academy of Medical Sciences, Jinan, China.
- 3Clinical Medical and Preclinical Medicine School, Shandong First Medical University & Shandong Academy of Medical Sciences, Jinan, China.
- 4College of Business, City University of Hong Kong, Hong Kong, Hong Kong SAR, China.
Abstract
Cross-subject electroencephalogram (EEG) recognition remains a fundamental challenge because large inter-subject variability causes severe distribution shifts that substantially degrade the generalization ability of existing learning models. Most existing transfer learning methods attempt to align global feature distributions but often overlook subject-specific structural variations, resulting in unstable adaptation under limited calibration data. To address this problem, we propose a Prototype-Consistent Style Transfer Mapping (PC-STM) framework for cross-subject neuropsychiatric disease recognition. Unlike conventional domain adaptation methods, the proposed framework models subject discrepancy as a prototype-guided style transformation that explicitly aligns target representations with class-consistent source prototypes while preserving discriminative structures. Specifically, a multi-source prototype construction strategy is introduced to characterize representative class distributions, followed by a confidence-aware transductive semi-supervised style transfer module that iteratively refines transformation parameters using labeled calibration samples and unlabeled evaluation samples. Furthermore, a reliability-aware ensemble learning strategy dynamically aggregates multiple subject-specific classifiers according to their calibration consistency, thereby improving robustness against heterogeneous subject distributions. Experiments on two public EEG datasets covering major depressive disorder and Parkinson's disease show stable performance under the evaluated leave-one-pair-out protocol. The calibration labels are inherited from the public dataset cohorts or task conditions, and the study does not constitute prospective clinical diagnostic validation. Beyond the evaluated EEG tasks, prototype-guided style transfer may provide a basis for future investigation of other limited-supervision recognition problems.
Related
- Mapping Habitat Suitability of the Asiatic Black Bear (Ursus thibetanus) in the Kalam Valley, Pakistan: An Ensemble Modelling Approach.
- Mapping of TEMPO-, PROXYL-, and Phosphoryl-Functionalized Spin Probes into Biointerface Fatty Acid Recognition Sites of BSA
- Multimodal Magnetic Resonance Imaging Biomarkers of Pediatric Traumatic Brain Injury Identified by Ensemble Learning Are Associated with Psychopathology
- Mangrove taxon mapping from Sentinel-1/2 using a compact one-dimensional network with probability calibration, predictive entropy, and frozen transfer evaluation
- A comparative analysis of machine learning, deep learning, and ensemble learning models for suicidal ideation detection on social media
BibTeX
@article{ding2026style,
title = {{Style transfer mapping with ensemble learning for cross-subject EEG-based neuropsychiatric disease recognition.}},
author = {Xinmin Ding and Yiming Guo and Shuwei Ge and Shenting Yang and Jipu Su and Minpeng Xu and Xuenan Wang and Hongli Chang},
journal = {Frontiers in physiology},
year = {2026},
month = oct,
doi = {10.3389/fphys.2026.1947840},
url = {https://europepmc.org/article/MED/42952503},
note = {CC BY},
}
RIS
TY - JOUR
TI - Style transfer mapping with ensemble learning for cross-subject EEG-based neuropsychiatric disease recognition.
AU - Xinmin Ding
AU - Yiming Guo
AU - Shuwei Ge
AU - Shenting Yang
AU - Jipu Su
AU - Minpeng Xu
AU - Xuenan Wang
AU - Hongli Chang
PY - 2026
DA - 2026/10/14
JO - Frontiers in physiology
DO - 10.3389/fphys.2026.1947840
UR - https://europepmc.org/article/MED/42952503
AB - Cross-subject electroencephalogram (EEG) recognition remains a fundamental challenge because large inter-subject variability causes severe distribution shifts that substantially degrade the generalization ability of existing learning models. Most existing transfer learning methods attempt to align global feature distributions but often overlook subject-specific structural variations, resulting in unstable adaptation under limited calibration data. To address this problem, we propose a Prototype-Consistent Style Transfer Mapping (PC-STM) framework for cross-subject neuropsychiatric disease recognition. Unlike conventional domain adaptation methods, the proposed framework models subject discrepancy as a prototype-guided style transformation that explicitly aligns target representations with class-consistent source prototypes while preserving discriminative structures. Specifically, a multi-source prototype construction strategy is introduced to characterize representative class distributions, followed by a confidence-aware transductive semi-supervised style transfer module that iteratively refines transformation parameters using labeled calibration samples and unlabeled evaluation samples. Furthermore, a reliability-aware ensemble learning strategy dynamically aggregates multiple subject-specific classifiers according to their calibration consistency, thereby improving robustness against heterogeneous subject distributions. Experiments on two public EEG datasets covering major depressive disorder and Parkinson's disease show stable performance under the evaluated leave-one-pair-out protocol. The calibration labels are inherited from the public dataset cohorts or task conditions, and the study does not constitute prospective clinical diagnostic validation. Beyond the evaluated EEG tasks, prototype-guided style transfer may provide a basis for future investigation of other limited-supervision recognition problems.
SN - 1664-042X
ER -