Automatic decomposition of electrophysiological data into distinct nonsinusoidal oscillatory modes

Neurophysiological signals are often noisy, nonsinusoidal, and consist of transient bursts. Extraction and analysis of oscillatory features (such as waveform shape and cross-frequency coupling) in such data sets remains difficult. This limits our understanding of brain dynamics and its functional importance. Here, we develop iterated masking empirical mode decomposition (itEMD), a method designed to decompose noisy and transient single-channel data into relevant oscillatory modes in a flexible, fully data-driven way without the need for manual tuning. Based on empirical mode decomposition (EMD), this technique can extract single-cycle waveform dynamics through phase-aligned instantaneous frequency. We test our method by extensive simulations across different noise, sparsity, and nonsinusoidality conditions. We find itEMD significantly improves the separation of data into distinct nonsinusoidal oscillatory components and robustly reproduces waveform shape across a wide range of relevant parameters. We further validate the technique on multimodal, multispecies electrophysiological data. Our itEMD extracts known rat hippocampal θ waveform asymmetry and identifies subject-specific human occipital α without any prior assumptions about the frequencies contained in the signal. Notably, it does so with significantly less mode mixing compared with existing EMD-based methods. By reducing mode mixing and simplifying interpretation of EMD results, itEMD will enable new analyses into functional roles of neural signals in behavior and disease. NEW & NOTEWORTHY We introduce a novel, data-driven method to identify oscillations in neural recordings. This approach is based on empirical mode decomposition and reduces mixing of components, one of its main problems. The technique is validated and compared with existing methods using simulations and real data. We show our method better extracts oscillations and their properties in highly noisy and nonsinusoidal datasets.