Neural oscillation and non-oscillation separator for highly time-resolved decomposition of beta band activity

Neural population activity exhibits rhythmic oscillations across multiple frequency bands, superimposed on a broadband, aperiodic background that follows a characteristic 1/f-like spectral structure. Accurate interpretation of time-varying neural dynamics therefore requires effective separation of periodic (oscillatory) and aperiodic (non-oscillatory) components. However, conventional bandpass filtering ignores aperiodic contributions, while recent decomposition approaches operate primarily in the frequency domain, limiting temporal resolution and hindering the detection of rapidly evolving oscillatory activity. Here, we introduce the Neural Oscillation and Non-Oscillation Separator (NONOS), a signal decomposition framework that enables time-resolved separation of periodic and aperiodic components in neural recordings. NONOS employs a deep neural network whose output is constrained to follow a 1/f-like spectral profile of aperiodic activity. By taking time-domain neural signals as input and directly decomposing periodic and aperiodic components in the time domain, this approach overcomes a fundamental limitation of frequency-domain-based methods. While the proposed approach is applicable across frequency bands, we demonstrate its utility using beta band activity as a representative and challenging case, given its rapid, state-dependent modulation in motor and subcortical circuits. Applying NONOS to local field potential recordings from patients with Parkinson's disease, we show that improved removal of the 1/f-like spectral component enhances the reliability and interpretability of behavior decoding based on beta band oscillations. Furthermore, in a hemiparkinsonian mouse model, NONOS sensitively captures stimulation-induced changes in subthalamic beta band activity with higher temporal fidelity than existing methods. Together, these results establish NONOS as a generalizable approach for time-resolved decomposition of neural signals.