SCALE-CONDITIONAL MOTIF DISCOVERY FOR MULTI-DIMENSIONAL TIME SERIES

Identifying recurrent patterns in high-dimensional time series is a central challenge in neuroscience, particularly for large-scale recordings that integrate multiple modalities and span diverse timescales.

A major limitation of existing motif discovery approaches is scale ambiguity: recurrent patterns often occur across multiplexed temporal scales, making their detection and interpretation challenging.

Here, we introduce a scale-conditional framework for unsupervised motif discovery that explicitly separates intrinsic timescales prior to pattern detection. By constraining motif identification to scale-specific signal components, this approach prevents motifs from collapsing across unrelated dynamics, substantially improving interpretability while reducing the effective search space.

Using simulations with known ground truth, we show that scale-conditional motif discovery enables more precise and robust identification of recurrent patterns across a wide range of noise levels, dimensionalities, and overlapping dynamics compared to conventional approaches. We further establish a statistical curation and filtering strategy that improves specificity and robustness in high-dimensional settings.

We demonstrate the generality of the framework across diverse neuroscience datasets, including multimodal recordings integrating behavior, neuromodulator dynamics, and electrophysiology, as well as population spiking activity, calcium imaging, and local field potential signals. This enables state- and behavior-specific motif analysis and provides a direct link between recurrent patterns and underlying population dynamics.

The framework is fully unsupervised, template-free and deterministic, with few and interpretable parameters. Together, this work introduces a principled approach for discovering interpretable, scale-specific motifs in complex neural data, facilitating the study of brain dynamics across timescales, modalities, and behavioral states.