Preventing Data Leakage in EEG-Based Survival Prediction: A Two-Stage Embedding and Transformer Framework

View PDF HTML (experimental)

Abstract:Deep learning models have shown promise in EEG-based outcome prediction for comatose patients after cardiac arrest, but their reliability is often compromised by subtle forms of data leakage. In particular, when long EEG recordings are segmented into short windows and reused across multiple training stages, models may implicitly encode and propagate label information, leading to overly optimistic validation performance and poor generalization. In this study, we identify a previously overlooked form of data leakage in multi-stage EEG modeling pipelines. We demonstrate that violating strict patient-level separation can significantly inflate validation metrics while causing substantial degradation on independent test data. To address this issue, we propose a leakage-aware two-stage framework. In the first stage, short EEG segments are transformed into embedding representations using a convolutional neural network with an ArcFace objective. In the second stage, a Transformer-based model aggregates these embeddings to produce patient-level predictions, with strict isolation between training cohorts to eliminate leakage pathways. Experiments on a large-scale EEG dataset of post-cardiac-arrest patients show that the proposed framework achieves stable and generalizable performance under clinically relevant constraints, particularly in maintaining high sensitivity at stringent specificity thresholds. These results highlight the importance of rigorous data partitioning and provide a practical solution for reliable EEG-based outcome prediction.

Comments: 9 pages, 2 figures. Preliminary version

Subjects:

Machine Learning (cs.LG)

Cite as: arXiv:2603.25923 [cs.LG]

(or arXiv:2603.25923v1 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2603.25923

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Yixin Zhou [view email] [v1] Thu, 26 Mar 2026 21:28:23 UTC (1,124 KB)