Time-series ECG Imputation Using a Pattern-Based Masking Framework

Area of Science:

• Biomedical Engineering
• Data Science
• Cardiology

Background:

• Continuous electrocardiogram (ECG) monitoring is crucial in hospitals, but missing data hinders real-time predictive systems.
• Existing research on ECG time-series imputation is limited, and current benchmarking often uses unrealistic random masking.

Purpose of the Study:

• To evaluate and benchmark various imputation methods for continuous ECG time-series data.
• To compare performance under both random and realistic pattern-based missingness conditions.

Main Methods:

• Extracted time-domain features from 40 patients' 12-lead Holter recordings (2.5-4 hours each).
• Introduced missingness using random and pattern-based masking.
• Compared seven imputation methods: global mean, linear interpolation, KNN, MICE, softImpute, SMILES, and SAITS.
• Evaluated performance using Mean Absolute Error (MAE).

Main Results:

• All imputation methods showed higher MAE under pattern-based masking compared to random masking.
• SAITS demonstrated the best performance across both masking types.
• Simpler methods like SoftImpute and KNN showed competitive performance, especially at certain missingness levels.

Conclusions:

• Random masking may underestimate the real-world accuracy of time-series imputation techniques.
• Context-specific imputation strategies, considering masking approach and method, are vital.
• Balancing model complexity with practical factors is essential for real-time ECG data deployment.