[Paper Review] High Frequency EEG Artifact Detection with Uncertainty via Early Exit Paradigm

Lorena Qendro, Alexander B. Campbell+2 more|arXiv (Cornell University)|Jul 21, 2021

EEG and Brain-Computer Interfaces4 citations

TL;DR

This paper proposes E4G, a deep learning framework for high-frequency EEG artifact detection that leverages the early exit paradigm to enable uncertainty-aware, single-forward-pass predictions. By forming an implicit ensemble of models through early exits, E4G achieves state-of-the-art F1 score (84%) and well-calibrated uncertainty—comparable to Monte Carlo dropout but with 5.2x lower latency—making it ideal for clinician-in-the-loop systems.

ABSTRACT

Electroencephalography (EEG) is crucial for the monitoring and diagnosis of brain disorders. However, EEG signals suffer from perturbations caused by non-cerebral artifacts limiting their efficacy. Current artifact detection pipelines are resource-hungry and rely heavily on hand-crafted features. Moreover, these pipelines are deterministic in nature, making them unable to capture predictive uncertainty. We propose E4G, a deep learning framework for high frequency EEG artifact detection. Our framework exploits the early exit paradigm, building an implicit ensemble of models capable of capturing uncertainty. We evaluate our approach on the Temple University Hospital EEG Artifact Corpus (v2.0) achieving state-of-the-art classification results. In addition, E4G provides well-calibrated uncertainty metrics comparable to sampling techniques like Monte Carlo dropout in just a single forward pass. E4G opens the door to uncertainty-aware artifact detection supporting clinicians-in-the-loop frameworks.

Motivation & Objective

• To address the limitations of deterministic, resource-heavy EEG artifact detection pipelines that rely on hand-crafted features.
• To enable per-time-point artifact detection with predictive uncertainty estimation for improved clinical trust and decision-making.
• To develop a scalable, efficient deep learning framework that provides uncertainty quantification without sampling-based methods like Monte Carlo dropout.
• To support clinician-in-the-loop systems by identifying high-uncertainty predictions for expert review.

Proposed method

• E4G employs the early exit (EE) paradigm in a deep neural network, where multiple intermediate exit branches generate predictions at different network depths.
• Each early exit branch produces a per-time-point prediction, forming an implicit ensemble of models that enables uncertainty estimation.
• The model is trained using a weighted sum of losses across all exit points, with learnable weights to prioritize early and late predictions.
• Predictive uncertainty is quantified using predictive entropy and Brier score, computed from the ensemble of exit predictions.
• The framework enables single-forward-pass uncertainty estimation, avoiding the computational cost of Monte Carlo dropout.
• The method supports uncertainty-informed decision making by analyzing disagreement across exit branches for artifact boundary detection and confidence assessment.

Experimental results

Research questions

• RQ1 Can the early exit paradigm be effectively applied to EEG artifact detection to enable uncertainty quantification without sampling-based methods?
• RQ2 How does E4G's uncertainty estimation compare to Monte Carlo dropout in terms of calibration and predictive performance?
• RQ3 Can E4G achieve state-of-the-art artifact detection performance while maintaining low inference latency?
• RQ4 How can predictive uncertainty from multiple early exits inform clinical decision-making in a clinician-in-the-loop framework?
• RQ5 Does the implicit ensemble formed by early exits improve robustness to false predictions compared to deterministic models?

Key findings

• E4G achieves a state-of-the-art F1 score of 84% on the TUH-A v2.0 test set, matching the performance of a vanilla baseline.
• E4G provides well-calibrated uncertainty with a predictive entropy of 0.45 and Brier score of 0.53, outperforming Monte Carlo dropout (0.53 and 0.55, respectively) on incorrect predictions.
• E4G incurs only 1.1x latency overhead compared to the baseline, while Monte Carlo dropout increases latency by 5.2x with 5 samples.
• Disagreement between early exit predictions reveals uncertainty in artifact boundaries, enabling automatic detection of ambiguous regions for clinician review.
• The framework enables uncertainty-aware decision making, such as routing high-uncertainty samples to experts, without sacrificing accuracy.
• The implicit ensemble formed by early exits supports interpretability and robustness, with exit diversity useful for future ensemble aggregation strategies.