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

2026-06-07 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Health & Medical Research · Depth: Expert, quick

Summary

A new spectral audit framework has been developed to analyze deep learning models processing physiological time series, specifically EEG and ECG data. This framework combines aperiodic/periodic decomposition, phase-preserving Fourier interventions, sham controls, and simulation validation to assess model reliance on the broadband aperiodic 1/f-like envelope, which covaries with arousal, age, and pathology. The audit revealed that aperiodic reliance is task-dependent and architecture-general across six neural architectures. For sleep-wake classification, flattening the aperiodic component led to drops exceeding 0.42 balanced-accuracy points, while clinical abnormality detection saw drops of 0.07-0.13. Motor imagery tasks showed minimal reliance. Furthermore, six of seven EEG foundation models exhibited FDR-significant aperiodic reliance on clinical EEG, an effect that persisted even after age/sex and recording-era controls. Applying the framework to PTB-XL ECG data confirmed this confound class extends beyond EEG, with neural drops of 0.32-0.36 after demographic matching.

Key takeaway

For machine learning engineers developing deep learning models for physiological time series like EEG or ECG, you must integrate aperiodic controls into your development and interpretation workflows. Ignoring the broadband aperiodic 1/f-like envelope can lead to misinterpretations of model performance and underlying mechanisms, especially for tasks like sleep-wake classification or clinical abnormality detection. Systematically audit your models' reliance on these signals to ensure robust and truly interpretable results.

Key insights

Deep learning models on physiological data often rely on aperiodic signals, impacting task performance.

Principles

• Aperiodic reliance is task-dependent.
• Aperiodic reliance is architecture-general.
• Aperiodic controls should be standard.

Method

A spectral audit framework combines aperiodic/periodic decomposition, phase-preserving Fourier interventions, sham controls, and simulation validation to assess model reliance.

In practice

• Implement aperiodic controls in physiological DL.
• Account for aperiodic signals in EEG/ECG models.
• Evaluate task-specific aperiodic reliance.

Topics

• Deep Learning
• Physiological Time Series
• EEG Analysis
• ECG Analysis
• Spectral Audit Framework
• Aperiodic Signals

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• SafeECGMatch: Calibration-Aware Joint Frequency and Time Space Semi-Supervised Learning for Open-Set ECG Classification — SafeECGMatch improves ECG classification by safely handling out-of-distribution data within semi-supervised learning.
• Pulse-Driven Neural Architecture: Learnable Oscillatory Dynamics for Robust Continuous-Time Sequence Processing — Biologically-inspired oscillatory dynamics significantly improve neural network robustness to temporal input gaps.
• A self-supervised electrocardiogram foundation model for empowering cardiovascular disease prediction and genetic factor discovery — A self-supervised ECG foundation model improves CVD prediction and uncovers novel genetic associations.
• Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis — CLMT uses a two-stage discrete translation to synthesize physiological signals across modalities and frequencies efficiently.
• ECG-Lens: Benchmarking ML & DL Models on PTB-XL Dataset — Deep learning models, especially complex CNNs, significantly outperform traditional ML for raw 12-lead ECG classification.
• PhysioMotion Artifact: A task-driven EEG dataset with point-wise motion artifact annotations — PhysioMotion Artifact is a new EEG dataset with detailed point-wise annotations for 14 types of physiological motion artifacts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.