Title: A Novel Explainable Hierarchical Self-Attention Framework for Time-Domain Tremor Detection

Abstract:

Tremors, a prevalent movement disorder linked to disorders such as Essential tremor and Parkinson’s disease, are conventionally diagnosed via expert clinical evaluation. While current automated detection systems often depend on frequency-domain features derived from clinical knowledge, this study introduces a novel, explainable two-stage hierarchical framework designed to detect tremors directly in the time domain. This approach learns tremor patterns straight from 3D kinematic marker time-series data collected throughout entire tremor-provoking trials.

The proposed architecture integrates a deep convolutional network with a long short-term memory (LSTM) model to extract representations from short, discrete, non-overlapping segments of kinematic data. These segment features are subsequently analyzed by a vision transformer, which captures long-term temporal dynamics to enable trial-level classification.

When tested across nine distinct body parts, the framework yielded F1-scores ranging from 0.594 to 0.947, with an average score of 0.765. Although these results lag behind the state-of-the-art performance of frequency-domain methods (0.909), the model requires significantly less preprocessing. Furthermore, the use of attention weights and gradient-based class activation maps (Grad-CAM) allowed for the identification of specific time-domain tremor features across different body regions. This proof-of-concept study highlights the viability of data-driven time-domain modeling for tremor detection across diverse anatomical locations, minimizing the need for expert-engineered spectral features while offering post-hoc interpretability of both temporal and anatomical tremor patterns.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC