Uncertainty quantification in DenseNet model using myocardial infarction ECG signals

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing
• Machine Learning for Diagnostics

Background:

• Myocardial infarction (MI) diagnosis relies on electrocardiograms (ECGs), but noise can impede automated analysis.
• Quantifying model uncertainty is crucial for reliable AI-driven ECG interpretation.

Purpose of the Study:

• Develop and evaluate a Dirichlet DenseNet model for robust MI detection from ECGs.
• Assess the model's ability to identify and quantify uncertainty in noisy ECG signals.
• Ensure trustworthy AI for critical healthcare applications like emergency MI diagnosis.

Main Methods:

• A Dirichlet DenseNet model was trained on ECGs from the PTB database.
• The model was tested using synthesized ECGs with added electromagnetic (em) and motion artifact (ma) noise.
• Predictive entropy served as the uncertainty measure, with performance evaluated using UNSE, UNSP, UNAC, and UNPR metrics.

Main Results:

• The model's uncertainty sensitivity (UNSE) increased with decreasing noise levels.
• High uncertainty accuracy (UNAC) was achieved for both em (80%) and ma (82.4%) noise at specific signal-to-noise ratios (SNRs).
• Uncertainty specificity (UNSP) and precision (UNPR) approached 100%, indicating strong self-awareness of prediction confidence.

Conclusions:

• The developed DenseNet model reliably conveys diagnostic uncertainty, enhancing trust in AI for ECG analysis.
• The model's ability to handle noisy data and communicate confidence makes it suitable for emergency MI diagnosis.
• This research supports the integration of AI tools in clinical settings for improved patient outcomes.