A spectral audit framework reveals task-dependent aperiodic reliance across EEG and ECG deep learning

2026-06-07

AI Summary

This study addresses a critical yet overlooked issue in deep learning models for physiological time series: their potential overreliance on broadband aperiodic (1/f-like) components rather than task-relevant periodic features, which may compromise interpretability and generalization. The authors propose a spectral auditing framework that integrates aperiodic/periodic component decomposition, phase-preserving Fourier-based interventions, pseudo-controls, and simulation-based validation to systematically assess model dependence on aperiodic signals. Their analysis reveals, for the first time, that such dependence is both task-specific and prevalent across diverse architectures, with significant effects observed in both EEG and ECG data. Experiments across sleep staging, clinical anomaly detection, and the PTB-XL ECG benchmark show performance drops of 0.07–0.42 points in most baseline models upon removal of aperiodic components, underscoring the necessity of incorporating aperiodic signal control into physiological modeling paradigms.

Abstract

Deep learning on physiological time series is interpreted through domain-specific features -- oscillatory rhythms in EEG, morphological complexes in ECG -- yet these signals sit atop a broadband aperiodic 1/f-like envelope that covaries with arousal, age, and pathology. We introduce a spectral audit framework combining aperiodic/periodic decomposition, phase-preserving Fourier interventions, sham controls, and simulation validation. Aperiodic reliance was task-dependent and architecture-general: across six neural architectures, flattening drops exceeded 0.42 balanced-accuracy points for sleep-wake classification, reached 0.07-0.13 for clinical abnormality detection, and remained minimal for motor imagery. Six of seven EEG foundation models showed FDR-significant aperiodic reliance on clinical EEG; age/sex and recording-era controls reduced but did not eliminate the effect. Applying the audit to PTB-XL ECG revealed neural drops of 0.32--0.36 persisting after demographic matching, confirming this confound class extends beyond EEG. Aperiodic controls should become standard for interpretable physiological time-series deep learning.

Problem

• aperiodic
• physiological time series
• deep learning
• spectral confound
• interpretability

Innovation

• spectral audit
• aperiodic signal
• physiological time series
• deep learning interpretability
• 1/f noise