Towards Dataset Agnostic ECG Signal Quality Assessments for Psychophysiological Analysis

Determining the relationship between psychological states and physiological processes requires accurate measurement of physiological signals associated with the autonomic nervous system’s activity, such as electrocardiograms. Electrocardiogram-derived measures can be associated with changes in emotional or affective states, but only if they are recorded with sufficient quality. Poor signal quality can distort the waveform and undermine the reliability of any downstream analyses, underscoring the need for robust signal quality assessment. There is currently no universally applicable electrocardiogram signal quality assessment adaptable across diverse psychophysiological datasets without extensive manual tuning. Current heuristic or machine learning-based approaches are tailored to individual datasets and require large training datasets, consisting of several hours of annotated electrocardiogram recordings containing real-world quality issues. Creating these datasets is slow and costly because domain experts must accurately annotate waveform distortions, so high-quality datasets remain scarce. This work contributes a novel approach that utilizes adaptive quality checks based on each recording’s intrinsic statistical properties to create a quality assessment that does not rely on dataset-specific tuning. While minor adjustments (e.g., correlation thresholds, lead-aggregation logic) can be made to align with different labeling standards, the approach remains robust and transferable across varied electrocardiogram signals. The limitations of existing methods are addressed by accurately assessing signal quality across varied recording conditions, devices, and populations. Results indicate near state-of-the-art performance on both controlled (91% accuracy) and ambulatory (96% accuracy) datasets, while remaining dataset agnostic. Assessments of the impact of signal quality on emotion classification reinforce that even subtle artefacts can destabilize heart rate variability features and impair model performance, highlighting the necessity of applying signal quality assessment prior to affective modeling.