1304-GRP CG-ECG: A Selective Classification Abstention Framework for Trustworthy ECG Classification
Location
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Document Type
Event
Start Date
20-11-2026 2:00 PM
Description
Reliable electrocardiogram (ECG) classification requires more than assigning every rhythm segment to the most likely class. In computer- aided diagnosis, a model should also know when a prediction is not sufficiently reliable for automatic retention and should instead be de- ferred to human-in-the-loop review. This paper presents CG-ECG, a two- gate confidence-geometry abstention framework for trustworthy ECG- based computer-aided diagnosis. The first gate evaluates output-level confidence using maximum softmax probability or predictive entropy. The second gate evaluates representation-level support by projecting penultimate-layer embeddings with singular value decomposition and constructing class-specific convex-hull regions from true-positive train- ing embeddings. A prediction is retained only when it passes a confi- dence gate and is geometrically supported by the predicted class region. The framework is evaluated on two public ECG benchmarks: a MIT- BIH arrhythmia classification setting using 1D ECG signals and 2D waveform-image representations and a PTB-XL 1D time-series bench- mark using a stratified held-out test split. Results show that confidence gates are strong selective baselines, achieving zero retained errors at approximately 60 to 64% coverage on MIT-BIH. Geometry alone does not uniformly outperform confidence thresholding, but it provides com- plementary representation-level evidence and auditable deferral causes. On PTB-XL, the geometry-only gate at k = 2 reduces retained er- rors by 50% relative to forced classification while preserving high cover- age. These findings position confidence-geometry abstention as an inter- pretable selective-classification framework for ECG-based CAD
1304-GRP CG-ECG: A Selective Classification Abstention Framework for Trustworthy ECG Classification
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Reliable electrocardiogram (ECG) classification requires more than assigning every rhythm segment to the most likely class. In computer- aided diagnosis, a model should also know when a prediction is not sufficiently reliable for automatic retention and should instead be de- ferred to human-in-the-loop review. This paper presents CG-ECG, a two- gate confidence-geometry abstention framework for trustworthy ECG- based computer-aided diagnosis. The first gate evaluates output-level confidence using maximum softmax probability or predictive entropy. The second gate evaluates representation-level support by projecting penultimate-layer embeddings with singular value decomposition and constructing class-specific convex-hull regions from true-positive train- ing embeddings. A prediction is retained only when it passes a confi- dence gate and is geometrically supported by the predicted class region. The framework is evaluated on two public ECG benchmarks: a MIT- BIH arrhythmia classification setting using 1D ECG signals and 2D waveform-image representations and a PTB-XL 1D time-series bench- mark using a stratified held-out test split. Results show that confidence gates are strong selective baselines, achieving zero retained errors at approximately 60 to 64% coverage on MIT-BIH. Geometry alone does not uniformly outperform confidence thresholding, but it provides com- plementary representation-level evidence and auditable deferral causes. On PTB-XL, the geometry-only gate at k = 2 reduces retained er- rors by 50% relative to forced classification while preserving high cover- age. These findings position confidence-geometry abstention as an inter- pretable selective-classification framework for ECG-based CAD