Researchers have developed a highly accurate, explainable machine learning model that can differentiate between essential tremor (ET) and cortical myoclonus (CM) using wearable sensor data.
Study: Explainable machine learning for movement disorders - Classification of tremor and myoclonus. Image Credit: Vitalii Vodolazskyi/Shutterstock.com
A new study published in Computers in Biology and Medicine outlines how this approach, based on accelerometry (ACC) data and power spectrum analysis, achieved near-perfect classification across a range of movement tasks. By analyzing motion patterns from 19 ET and 19 CM patients during 21 tasks, the model—using spectral features in the 3–10 hertz (Hz) range—offers a promising step toward more objective, data-driven diagnostics for these often-overlapping movement disorders.
Background
Hyperkinetic disorders like ET and CM are notoriously difficult to distinguish clinically due to overlapping symptoms, particularly in the limbs. Diagnosis often depends on subjective judgment and traditional electrophysiological tools, with little in the way of standardized, evidence-based criteria.
While machine learning has been applied to tremor analysis in contexts like detection and subtyping, no prior studies have directly addressed differentiating ET from CM. Previous ML approaches vary widely in preprocessing and modeling strategies, often trading interpretability for performance.
This study fills that gap by applying generalized matrix learning vector quantization (GMLVQ), an explainable ML method, to ACC-derived power spectra. The approach preserves clinically meaningful frequency-domain patterns without relying on arbitrary feature engineering. By identifying task- and sensor-specific spectral signatures, it offers a clear, interpretable framework to support diagnosis and guide targeted treatment strategies.
Study Design and Methods
Participants performed 21 upper-limb tasks while wearing eight ACC sensors. The data was transformed into power spectra (1–30 Hz) and analyzed using GMLVQ, which identifies both class prototypes and frequency relevance. Tasks included resting, postural, and dynamic conditions, with sensor data aggregated by limb or averaged bilaterally.
Classification performance was especially strong in the 3–4 Hz, 5–7 Hz, and 9–10 Hz ranges—frequencies associated with tremor and myoclonus patterns. Dynamic tasks like finger-to-nose and distal sensors (e.g., on the hands and fingers) delivered the highest accuracy. Importantly, the model's explainability revealed clear contrasts: ET exhibited rhythmic peaks, while CM showed broader, more variable frequency patterns. The method’s robustness was confirmed by comparing it with random forest models.
Key Findings
With ACC data from 19 ET and 19 CM patients (validated by clinical experts), the model achieved exceptional classification across 227 task-sensor combinations. Postural tasks were particularly effective, with pronated outstretched arms reaching perfect discrimination (AUROC = 1.0). Dynamic tasks like finger tapping also performed well, while resting conditions showed weaker results (max AUROC = 0.76). Sensor setups spanning multiple body regions proved most effective, with a median AUROC of 0.96.
The model’s explainability provided key insights. ET presented narrow 5–7 Hz tremor peaks flanked by frequency dips at 3–4 Hz and 9–10 Hz. In contrast, CM signals displayed broader spectral distributions. A two-dimensional projection of the model's output showed nearly perfect class separation, and the relevance matrix highlighted the most critical frequency interactions, especially around tremor-related peaks.
These findings validate the power of ML-driven spectral analysis to objectively distinguish ET from CM and support clinical decision-making with transparent, physiologically grounded outputs.
Conclusion
This study highlights how explainable ML, specifically GMLVQ applied to ACC data, can accurately and transparently differentiate between essential tremor and cortical myoclonus. By focusing on clinically meaningful frequency patterns and emphasizing task-specific performance, the method aligns with known pathophysiological features and offers an objective, interpretable diagnostic aid.
The model’s transparency—through frequency relevance matrices and class prototypes—supports clinician trust and enhances diagnostic workflows. These results offer a solid foundation for future research, including validation in larger, more diverse cohorts and integration with electrophysiological data. Ultimately, this approach shows strong potential for improving diagnostic precision and enabling more personalized treatment strategies in clinical neurology.
Journal Reference
van den Brandhof et al. (2025). Explainable machine learning for movement disorders - Classification of tremor and myoclonus. Computers in Biology and Medicine, 192, 110180. DOI: 10.1016/j.compbiomed.2025.110180. https://www-sciencedirect-com.proxy-ub.rug.nl/science/article/pii/S0010482525005311
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.