UCTECG-Net: Uncertainty-aware Convolution Transformer ECG Network for Arrhythmia Detection

Hamzeh Asgharnezhad; Pegah Tabarisaadi; Abbas Khosravi; Roohallah Alizadehsani; U. Rajendra Acharya

cs.LGFeb 18, 2026

UCTECG-Net: Uncertainty-aware Convolution Transformer ECG Network for Arrhythmia Detection

Authors: Hamzeh Asgharnezhad, Pegah Tabarisaadi, Abbas Khosravi, Roohallah Alizadehsani, U. Rajendra Acharya

Organizations: Deakin University, Australia . · School of Mathematics, Physics and Computing, University of Southern Queensland, Toowoomba, QLD, Australia .

Abstract

Deep learning has improved automated electrocardiogram (ECG) classification, but limited insight into prediction reliability hinders its use in safety-critical settings. This paper proposes UCTECG-Net, an uncertainty-aware hybrid architecture that combines one-dimensional convolutions and Transformer encoders to process raw ECG signals and their spectrograms jointly. Evaluated on the MIT-BIH Arrhythmia and PTB Diagnostic datasets, UCTECG-Net outperforms LSTM, CNN1D, and Transformer baselines in terms of accuracy, precision, recall and F1 score, achieving up to 98.58% accuracy on MIT-BIH and 99.14% on PTB. To assess predictive reliability, we integrate three uncertainty quantification methods (Monte Carlo Dropout, Deep Ensembles, and Ensemble Monte Carlo Dropout) into all models and analyze their behavior using an uncertainty-aware confusion matrix and derived metrics. The results show that UCTECG-Net, particularly with Ensemble or EMCD, provides more reliable and better-aligned uncertainty estimates than competing architectures, offering a stronger basis for risk-aware ECG decision support.

Figures & tables

Figure 1: Density plots of certainty values for correct and incorrect predictions, illustrating the model’s ability to separate confident correct outputs from uncertain errors.

		Correctness
		Correct	Incorrect
Confidence	Certain	CC	IC
Confidence	Uncertain	CU	IU

Table 1: Uncertainty confusion matrix

Panel of Figure 2 — Figure 2: Architecture of the LSTM-based model

Panel of Figure 3 — Figure 3: Architecture of the 1D CNN model

Panel of Figure 4 — Figure 4: Architecture of the Transformer model

Panel of Figure 5 — Figure 5: Two samples of normal and abnormal ECG signals and their corresponding spectrograms. Subfigures 5(a) and 5(b) display time-domain ECG waveforms for normal and abnormal cardiac activity, respectively. Subfigures 5(c) and 5(d) illustrate their time–frequency representations, showing clear spectral differences between normal and abnormal patterns.

Figure 6: Architecture of the proposed UCTECG-Net

Research	Dataset	Year	Approach	Metrics	UAcc
Evaluation of uncertainty quantification methods in multi-label classification Barandas et al. (2024)	CPSC, G12EC, PTB-XL	2024	BNN-Dropout, BNN-Laplace, Deep Ensemble, Bootstrap	AUROC, F1 Score, ECE	Reported
Fatigue Detection with Multimodal Physiological Signals via Uncertainty-Aware Deep Transfer Learning Kakhi et al. (2026)	Multimodal physiological dataset (EEG, ECG, EDA, PPG, RESP)	2026	Deep transfer learning MC-Dropout, Deep Ensembles, EMCD	Accuracy, F1 Score, Precision, Recall, AUC	Reported
Uncertainty quantification in DenseNet model using myocardial infarction ECG signals Jahmunah et al. (2023)	PTB (MI subset)	2023	Dirichlet DenseNet	Confusion Matrix, Uncertainty Confusion Matrix	Reported
A Reliable Deep Learning Model for ECG Interpretation: Mitigating Overconfidence and Direct UQ Li et al. (2025)	PhysioNet Challenge 2017	2025	Residual CNN, Dirichlet-based uncertainty, Subjective Logic	Accuracy, F1 Score, PR-AUC, ROC-AUC	Not reported
A novel inference system for detecting cardiac arrhythmia using deep learning framework Sai et al. (2025)	MIT-BIH	2025	DeepBiLSTMnet with wavelet sequence layer, Bi-LSTM	Accuracy, F1 Score, Precision, Recall	Not reported
This Research UCTECG-Net	MIT-BIH and PTB Arrhythmia	-	Multimodal Spectrum/Signal CNN, Transformer MC-Dropout, Deep Ensembles, EMCD	Accuracy, Sensitivity, Specificity, F1 Score	Reported

Table 2: Previous related studies summary

Dataset	MIT-BIH	PTB
Total Samples	109,446	14,552
Train Samples	87,554 (80%)	11,641 (80%)
Total Samples	21,892 (20%)	2,911 (20%)
Resampled Frequency	125 Hz	125 Hz
Time Steps 187	187	187
Number of Classes	5	2

Table 3: Datasets information

Model	Model	Accuracy	Precision	Recall	F1 Score
	LSTM	96.16 ± 2.92	95.57 ± 3.74	96.16 ± 2.92	95.56 ± 3.74
MIT-BIH	CNN1D	97.87 ± 0.08	97.79 ± 0.09	97.87 ± 0.08	97.77 ± 0.08
	Transformer	98.53 ± 0.02	98.50 ± 0.02	98.53 ± 0.02	98.49 ± 0.02
	UCTECG-Net	98.58 ± 0.02	98.54 ± 0.02	98.58 ± 0.02	98.54 ± 0.02
	LSTM	79.57 ± 4.12	78.91 ± 5.11	79.57 ± 4.12	78.16 ± 6.83
PTB	CNN1D	98.65 ± 0.32	98.65 ± 0.32	98.65 ± 0.32	98.65 ± 0.32

Table 4: Comparison of classification performance for Transformer, CNN1D, and LSTM models vs proposed UCTECG-Net. Results are reported as mean ± standard deviation over 5 runs.

Panel of Figure 7 — Figure 7: Confusion matrices of the best model, UCTECG-Net, for MIT-BIH (Subfigure 7(a) ) and PTB (Subfigure 7(b) ).

Dataset	Model	Method	CU	IC	IU	CC	UAcc	USen	USpe	UPre
	LSTM	MCD	998	926	2719	17249	83.35	51.87	86.38	26.85
	LSTM	Ensemble	315	142	1862	19573	90.85	68.93	91.31	14.47
	LSTM	EMCD	329	128	2012	19423	90.22	71.99	90.61	14.05
	CNN1D	MCD	439	205	1139	20109	93.86	68.17	94.64	27.82
MIT-BIH	CNN1D	Ensemble	190	267	293	21142	97.44	41.58	98.63	39.34
	CNN1D	EMCD	434	224	870	20364	95.0	65.96	95.9	33.28

Table 5: Comparison of uncertainty quantification results for the Transformer model using MCD, Ensemble, and EMCD methods.

Panel of Figure 8 — Figure 8: Subfigures 8(a) shows Uncertainty Accuracy per class for MIT-BIH. Subfigures 8(b) shows Uncertainty Accuracy per class for PTB

Electrocardiogram Classification

Electrocardiogram classification uses machine learning to identify heart rhythms and cardiac conditions from ECG waveforms for screening, diagnosis, and monitoring.

See all 106 papers

Robust Transfer Learning for Paper ECG Recognition
Sep 30, 2026Yinghao Xie, Zhenbang Dai, Haojun Wang +6Contrastive Learning Electrocardiography
Uppsala University, Sweden · Nanjing Medical University, China · The First Affiliated Hospital with Nanjing Medical University, China
TRACE: Expert-Aligned ECG Representation Learning with Rigorous Benchmarking and Real-World Validation in Acute Cardiac Care
Sep 28, 2026Lovely Yeswanth Panchumarthi, Andrew Lu, Saurabh Kataria +11Representation Learning Cross-Modal Representation Learning
Department of Computer Science, Emory University, GA, USA · School of Nursing, Emory University, GA, USA · Institute for Medical Engineering and Science, Massachusetts Institute of Technology, MA, USA +3
MambaCapsule: Towards Transparent Cardiac Disease Diagnosis with Electrocardiography Using Mamba Capsule Network
Jul 30, 2024Yinlong Xu, Zitai Kong, Yixuan Wu +5Interpretable ML Electrocardiogram Classification
State Key Laboratory of Transvascular Implantation Devices and TIDRI, Zhejiang University, Hangzhou, China · School of Public Health, Zhejiang University and Zhejiang Key Laboratory of Medical Imaging Artificial Intelligence, Hangzhu, China · State Key Laboratory of Transvascular Implantation Devices of The Second Affiliated Hospital and Liangzhu Laboratory, Zhejiang University, Hangzhou, China +3

Sep 30, 2026J/K move · Enter open · S save

Robust Transfer Learning for Paper ECG Recognition

Yinghao Xie, Zhenbang Dai, Haojun Wang, Jinyu Cai+5

Uppsala University, Sweden · Nanjing Medical University, China · The First Affiliated Hospital with Nanjing Medical University, China

Contrastive Learning Electrocardiography Electrocardiogram Classification Electrocardiogram Representation Learning

Open paper AskPDF arXiv

Paper ECG recognition is challenging because real-world ECG images vary in layout, physical artifacts, and label availability. We introduce RobECG-CL, a rank-aware contrastive learning framework for robust paper ECG representation learning. Starting from standard 12-lead ECG recordings, we construct progressively degraded paper ECG views with heterogeneous layouts and train the model to balance same-recording invariance with degradation-aware ordering. Across synthetic stress tests on CODE-II and EchoNext, RobECG-CL improves robustness under severe degradation and few-shot transfer, outperforming contrastive learning baselines and surpassing the waveform-based foundation model, ECG-FM, in the 1% labeled setting. On 312 samples of hospital data with 37 labels, RobECG-CL achieves the best macro AUROC.

Abstract

Electrocardiogram Classification

Explore similar work

Robust Transfer Learning for Paper ECG Recognition

TRACE: Expert-Aligned ECG Representation Learning with Rigorous Benchmarking and Real-World Validation in Acute Cardiac Care

MambaCapsule: Towards Transparent Cardiac Disease Diagnosis with Electrocardiography Using Mamba Capsule Network