Multi-dimensional signal exploration on the effect of tremor interventions

Background Machine learning in tremor time-series analysis has so far been applied to differentiate tremor disorders. Respective analytical approaches can similarly be geared to explore therapeutic interventions, which conventionally are assessed by their induced amplitude changes. This may overlook subtle signal alterations. Feature-based machine learning offers the potential to explore much more granular tremor signal changes induced by therapeutic interventions. Objective To investigate multidimensional time-series signal changes associated with tremor-modulating interventions, and to identify shared and distinct intervention-specific patterns across tremor conditions. Methods Sixty-six participants (Essential Tremor (ET): propranolol n = 13, deep brain stimulation (DBS) n = 10; Parkinson’s Disease (PD): levodopa n = 23; enhanced Physiological Tremor (ePT): caffeine n = 20) were assessed in therapeutic OFF and ON states using high-resolution tri-axial accelerometry. In addition to established tremor characteristics, advanced time-series features were extracted from raw recordings. Machine Learning (ML) approaches, including ensemble classification and Principal Component Analysis (PCA), were applied to detect intervention-related changes. Results Significant clinical and signal-based intervention effects were observed across cohorts with the strongest tremor suppression following DBS in ET and expected tremor enhancement following caffeine in ePT. While amplitude-related features decreased consistently after therapeutic interventions, dopaminergic treatment in PD was additionally associated with changes in frequency-related characteristics. Caffeine-induced tremor modulation formed a distinct multidimensional signal pattern, whereas therapeutic interventions showed partially overlapping effects. No single feature consistently characterized successful tremor modulation across all interventions. Conclusion Feature-based ML reveals intervention-specific tremor patterns beyond conventional measures, supporting data-driven approaches for objective tremor characterization.