MELEP: A Novel Predictive Measure of Transferability in Multi-label ECG Diagnosis

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiography (ECG) interpretation faces challenges due to limited annotated data.
• Transfer learning is crucial for ECG analysis, but assessing transferability remains underdeveloped.
• Existing methods lack specific metrics for multi-label ECG classification.

Purpose of the Study:

• To introduce MELEP (Muti-label Expected Log of Empirical Predictions), a novel metric for evaluating knowledge transfer effectiveness.
• To provide a computationally efficient and generic measure for pre-trained models in multi-label ECG diagnosis.
• To address the gap in transferability assessment for deep learning models in ECG interpretation.

Main Methods:

• Developed MELEP, a metric requiring a single forward pass of the pre-trained model.
• Tested MELEP's efficacy on diverse downstream multi-label ECG diagnosis tasks.
• Evaluated MELEP's correlation with actual model performance using fine-tuned deep neural networks (CNNs and RNNs).

Main Results:

• MELEP demonstrated strong predictive power for the performance of pre-trained models on small, imbalanced ECG datasets.
• High correlation coefficients (absolute values > 0.6) were observed between MELEP scores and average F1 scores.
• MELEP proved effective across different deep learning architectures (convolutional and recurrent networks).

Conclusions:

• MELEP is the first transferability metric specifically designed for multi-label ECG classification.
• MELEP can significantly expedite the selection of appropriate pre-trained models for ECG diagnosis.
• This metric reduces the need for extensive fine-tuning, saving computational resources and time.