Phenotypical differentiation of tremor using time series feature extraction and machine learning

Verena Häring, Veronika Selzam, Juan Francisco Martin‐Rodríguez, Petra Schwingenschuh, Gertrúd Tamás, 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. SchreglmannmedRxiv 2024 · DOI

TL;DR

Distinguishing Parkinson’s Disease from Essential Tremor is a notoriously difficult clinical task often prone to subjective bias. This study leverages "massive" time-series feature extraction and machine learning on accelerometer data from 370 patients to create a highly accurate, center-independent diagnostic tool. Beyond classification, the model provides a mechanistic "aha moment": Parkinsonian tremors fluctuate between multiple stable signal states (suggesting multiple oscillators), while Essential Tremor behaves like a singular pacemaker.

What Problem Does This Paper Address?

Tremor is the most common movement disorder, but telling the difference between Essential Tremor (ET) and tremor-dominant Parkinson’s Disease (PD) is surprisingly hard. Currently, doctors rely on subjective clinical observations, leading to misdiagnosis in over 30% of cases.

While researchers have developed objective sensor-based metrics—like the Tremor Stability Index (TSI)—these often fail to generalize. A metric that works in a lab in London might fail in a clinic in Budapest due to differences in equipment, recording positions, or patient populations. This paper seeks a "hypothesis-free" approach that can cut through this noise and provide a reliable, automated way to differentiate these disorders across multiple centers.

Methods

The researchers moved away from "hand-picked" features and instead used a "massive" feature extraction approach.

  1. Massive Feature Extraction: Using the hctsa (highly comparative time-series analysis) library, they extracted 7,873 mathematical descriptors from every 15-second accelerometer recording. These features cover everything from basic entropy and autocorrelation to complex non-linear wave dynamics.
  2. Machine Learning Pipeline: They tested several classifiers (Random Forest, kNN, Neural Nets), finding that Linear Support Vector Machines (SVM) performed best.
  3. Optimization: They systematically tested recording conditions. A key finding was that tri-axial data (calculating the vector amplitude sum of X, Y, and Z axes) outperformed mono-axial data by roughly 13-16% in accuracy.
  4. Feature Selection: Through univariate and multivariate testing, they whittled the thousands of features down to a "power duo" of two specific mathematical features that provided the best balance of accuracy and generalizability.

Datasets

The study utilized a large, multi-center dataset to ensure the results weren't just "overfitted" to one hospital's quirks:

  • Exploratory Dataset: 330 patients (158 ET, 172 PD) recruited from academic centers in London, Graz, Budapest, and Kiel.
  • Validation Dataset: A completely independent set of 40 patients (9 ET, 31 PD) from Nijmegen to test the model's real-world performance.
  • Total: 370 patients, making this one of the largest multi-center tremor accelerometry investigations to date.

Key Insights and Ideas

  • The Failure of "Classic" Metrics: Established metrics like TSI and peak power showed significant differences in the combined cohort but were inconsistent when looking at individual centers. This explains why previous mono-centric studies often fail to replicate.
  • Signal State Switching: One of the top-performing features (MF_hmm_CompareNStates) uses Hidden Markov Models to see if a signal stays in one "state" or jumps between several. PD tremors frequently switched between multiple stable states, while ET tremors remained in a singular state.
  • The "Pill-Rolling" Signature: Another top feature (MF_GARCHfit) measures signal skewness. PD signals were significantly more asymmetrical. The authors suggest this is the mathematical signature of the "pill-rolling" motion or sudden fluctuations characteristic of Parkinsonian rest tremor.
  • Rest is Best: Across all tests, recordings taken while the hand was at rest provided much better diagnostic separation than recordings taken during specific postures.

Key Results

  • Superior Accuracy: The ML model achieved 81.8% accuracy in differentiating ET from PD, significantly outperforming the best classic metric, TSI (70.4%).
  • High Sensitivity: The model showed 86.4% sensitivity for detecting the correct disorder.
  • Position Detection: The system was nearly perfect at distinguishing whether a patient was at rest or in a posture, regardless of their diagnosis (99.6% accuracy).
  • Independent Validation: When tested on the "unseen" Nijmegen dataset, the model maintained a solid 72.5% accuracy, proving it can handle data from a center it wasn't trained on.

Limitations and Future Work

The authors acknowledge that their dataset is predominantly of European-Caucasian descent, which may limit global generalizability. Additionally, the study did not include the recently defined "ET-plus" category (ET with subtle additional neurological signs), as the data was collected before that classification was standardized.

Future work aims to incorporate even larger, more diverse global datasets and explore whether these movement "signatures" can predict how well a patient will respond to specific therapies, such as deep brain stimulation.

Summary updated

Abstract

Abstract The reliable differentiation of tremor disorders poses a significant challenge, largely depending on the subjective interpretation of subtle signs and symptoms. Given the absence of a universally accepted bio-marker, diagnostic differentiation between the most prevalent tremor disorders, Essential Tremor (ET) and tremor-dominant Parkinson’s Disease (PD), frequently proves to be a non-trivial task. To address this, we employed massive time series feature extraction, a powerful tool to examine the entirety of mathematical descriptors of oscillating biological signals without imposing bias, in combination with machine-learning (ML). We applied this approach to accelerometer recordings from tremor patients to identify the optimal recording conditions, processing, and analysis settings, to differentiate ET and PD. We utilized hand accelerometer recordings from 370 patients (167 ET, 203 PD), clinically diagnosed at five academic centres specialising in movement disorders, comprising an exploratory (158 ET, 172 PD from London, Graz, Budapest, Kiel) and a validation dataset (9 ET, 31 PD from Nijmegen). Using 15 second recording segments from the more affected hand, we first extracted established, standardized tremor characteristics and assessed their cross-centre accuracy and validity. Second, we applied supervised ML to massive higher-order feature extraction of the same recordings to achieve optimal stratification and mechanistic exploration. While classic tremor characteristics were unable to consistently differentiate between conditions across centres, the resulting best classifying feature combination validated successfully. In comparison to tremor-stability index (TSI), the best performing classic tremor characteristic, feature-based analysis provided better classification accuracy (81.8% vs. 70.4%), sensitivity (86.4% vs. 70.8%) and specificity (76.6% vs. 70.2%), substantially improving stratification between ET and PD tremor. Similarly, this approach allowed the differentiation of rest from posture recordings independent of tremor diagnosis, again outperforming TSI (classification accuracy 99.6% vs. 49.2%). The interpretation of identified features indicates fundamentally different dynamics in tremor generating circuits: while there is an interaction between several central oscillators in the generation of PD rest tremor, resulting in several discrete but stable signal states, signal characteristics point towards a singular pacemaker in ET. This study highlights the limitations of current, established tremor metrics and establishes the use of feature-based machine learning as a powerful method to explore accelerometry-derived movement characteristics. More importantly, it showcases the strength of the combination of hypothesis-free, data-driven analyses and a large, multi-centre dataset. The results generated are thus resistant to device-, centre- and clinician-dependent bias and establish a generalizable differentiation method, representing a relevant step towards big data analysis in tremor disorders.
BibTeX
@misc{haring2024phenotypical,
  author = {Verena Häring and Veronika Selzam and Juan Francisco Martin‐Rodríguez and Petra Schwingenschuh and Gertrúd Tamás 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}},
  year = {2024},
  doi = {10.1101/2024.03.14.24303988},
  howpublished = {medRxiv},
}
Figure