Phenotypical Differentiation of Tremor Using Time Series Feature Extraction and Machine Learning

Verena Häring, Veronika Selzam, Juan Francisco Martin‐Rodríguez, Petra Schwingenschuh, Gertrúd Tamás, Linda Köhler, Jan Raethjen, Steffen Paschen, Franziska Goltz, Eoin Mulroy, Anna Latorre, Pablo Mir, Rick C. Helmich, Kailash P. Bhatia, Jens Volkmann, Robert L. Peach, Sebastian R. SchreglmannMovement 2025 · DOI

TL;DR

This paper introduces a novel machine learning approach that leverages extensive time-series feature extraction from accelerometer data to reliably differentiate between Essential Tremor (ET) and Parkinson's Disease (PD). By moving beyond traditional, less generalizable tremor metrics, this method offers a more accurate and mechanistic understanding of these common movement disorders.

What Problem Does This Paper Address?

Diagnosing tremor disorders, particularly distinguishing between Essential Tremor (ET) and Parkinson's Disease (PD), is a significant clinical challenge. Current diagnostic methods rely heavily on the interpretation of subtle clinical signs and symptoms, with no universally accepted biomarker available. This often leads to misdiagnoses and treatment difficulties. Existing quantitative methods, like the Tremor Stability Index (TSI) and mean harmonic power, have shown limitations in their generalizability across different clinical centers and patient cohorts. The authors aim to develop a more robust and generalizable method for differentiating ET and PD by employing a data-driven, hypothesis-free approach using machine learning on accelerometer data.

Methods

The core of this study lies in its innovative application of machine learning to a large, multicenter dataset of accelerometer recordings.

  • Massive Feature Extraction: Instead of relying on pre-defined tremor characteristics, the researchers employed the hctsa (highly comparative time-series analysis) algorithm to extract a vast number (7729) of mathematical features from the raw accelerometer data. This approach allows for an unbiased exploration of the time-series signal's complexity.
  • Supervised Machine Learning: Various supervised ML algorithms were trained and evaluated on this feature-rich dataset. The goal was to identify the optimal combination of features that could accurately classify ET and PD.
  • Model Evaluation and Validation: The study utilized a train/test split and 10-fold cross-validation to ensure robust model performance. Crucially, the findings were validated on independent datasets from different centers, assessing generalizability.
  • Feature Interpretation: Beyond classification, the authors delved into the mathematical underpinnings of the most effective features to gain mechanistic insights into the underlying tremor dynamics of ET and PD.

Datasets

The study utilized a substantial dataset comprising accelerometer recordings from 414 patients diagnosed with either ET or PD. These recordings were collected from six academic centers with expertise in movement disorders.

  • Exploratory Dataset: 158 ET and 172 PD patients.
  • Validation Dataset: 30 ET and 54 PD patients.

The data was collected from both rest and postural recordings, with a focus on the more severely affected hand to capture richer mechanical information. The multicenter nature of the dataset is a key strength, allowing for the assessment of generalizability across different clinical settings and recording protocols.

Key Insights and Ideas

  • Beyond Traditional Metrics: The study definitively shows that established tremor characteristics (like TSI) are insufficient for reliable, generalizable differentiation between ET and PD across multiple centers.
  • The Power of "Big Data" in Tremor: By combining a large, multicenter dataset with hypothesis-free feature extraction and ML, the researchers unlocked previously unrecognized disease-specific signal characteristics.
  • Mechanistic Clues from Features: The identified key features offer intriguing insights into the underlying pathophysiology. One feature suggests ET might involve a singular, stable "pacemaker," while PD exhibits multiple, discrete, and stable "signal states," hinting at different oscillatory circuit dynamics.
  • Triaxial Superiority: The study reinforces the importance of using triaxial accelerometers over monoaxial ones for capturing the full dimensionality of tremor signals, significantly impacting classification accuracy.

Key Results

  • Improved Classification Accuracy: The ML-based approach, using a combination of two key features (MF_GARCHfit and MF_hmm_CompareNStates), achieved a classification accuracy of 81.8% on the exploratory dataset, significantly outperforming traditional metrics like TSI (70.4%) and peak power (68.9%).
  • Validation Success: This combined ML model was successfully validated on independent datasets, achieving accuracies of 72.5% and 72.7% for the Nijmegen and Wuerzburg validation cohorts, respectively.
  • Feature Performance: The best-performing individual feature for univariate classification was MF_GARCHfit. However, the combination of MF_GARCHfit and MF_hmm_CompareNStates yielded the highest overall accuracy.
  • Rest vs. Postural Tremor: Similar to previous findings, rest tremor recordings generally provided more effective differentiation between ET and PD compared to postural recordings.

Limitations and Future Work

  • Generalizability: While the study used a multicenter dataset, the authors acknowledge that the population was predominantly of European-Caucasian descent. Future studies should incorporate larger, more diverse samples, including various genetic backgrounds, to establish truly generalizable tremor classification.
  • Data Curation: The authors emphasize the need for fully curated datasets, as outliers can disproportionately influence ML results.
  • Clinical Context: The ML-based approach focuses solely on movement characteristics from accelerometer data, not incorporating additional clinical information, which is crucial in daily practice.
  • Future Directions: The authors advocate for future ML-based studies to include larger and more diverse datasets, potentially through international collaboration, to further refine and validate these methods. They also suggest that the identified features could pave the way for more objective and potentially therapeutic interventions tailored to specific movement characteristics rather than just clinical diagnoses.

Summary updated

Abstract

This study establishes the use of feature-based ML as a powerful method to explore accelerometry-derived tremor signals. The combination of hypothesis-free, data-driven analyses and a large, multicenter dataset represents a relevant step towards big data analysis in tremor disorders. © 2025 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.
BibTeX
@inproceedings{haring2025phenotypical,
  author = {Verena Häring and Veronika Selzam and Juan Francisco Martin‐Rodríguez and Petra Schwingenschuh and Gertrúd Tamás and Linda Köhler and Jan Raethjen and Steffen Paschen and Franziska Goltz and Eoin Mulroy and Anna Latorre and Pablo Mir and Rick C. Helmich and Kailash P. Bhatia and Jens Volkmann and Robert L. Peach and Sebastian R. Schreglmann},
  title = {{Phenotypical Differentiation of Tremor Using Time Series Feature Extraction and Machine Learning}},
  booktitle = {Movement},
  year = {2025},
  doi = {10.1002/mds.70032},
}
Figure