Uncovering Signatures in Breath Turbulence for Data-Driven Identification of Individuals

Exhaled breath flow is turbulent, multiscale and uniquely influenced by the geometry of one's upper airway. These flows contain rich spatial and temporal structure such as vortical patterns, resonance modes, and spectral asymmetries that encode individual-specific information. In this work, we uncover interpretable features from breath velocity time series data that persist across recordings and subjects, revealing physically grounded patterns in breath turbulence.

From the breath velocity time series, we extract a set of interpretable and robust features, including multifractal descriptors (e.g., singularity strength and spectral asymmetry), energy localization, and long-range temporal correlations. These features, derived from both the time domain and time frequency representations, persist across recordings and allow strong separation between individuals.

Leveraging these insights, we build a breath-based identification framework that avoids retraining by representing users in a learned embedding space and computing similarity scores, enabling scalable comparisons across large populations. Our results demonstrate that interpretable, flow-informed features from breath turbulence can serve as reliable biometric markers. This work paves the way for developing non-invasive, physiology-based identification systems grounded in fluid dynamics and machine learning