CoChat

PHTNet: Characterization and Deep Mining of Involuntary Pathological Hand Tremor using Recurrent Neural Network Models

Open access

The DOI matches its Crossref record; 74 of the 74 references it cites were found in a registry or an index; 11 of the 28 sources its claims cite were checked for support.

Abstract

The global aging phenomenon has increased the number of individuals with age-related neurological movement disorders including Parkinson's Disease (PD) and Essential Tremor (ET). Pathological Hand Tremor (PHT), which is considered among the most common motor symptoms of such disorders, can severely affect patients' independence and quality of life. To develop advanced rehabilitation and assistive technologies, accurate estimation/prediction of nonstationary PHT is critical, however, the required level of accuracy has not yet been achieved. The lack of sizable datasets and generalizable modeling techniques that can fully represent the spectrotemporal characteristics of PHT have been a critical bottleneck in attaining this goal. This paper addresses this unmet need through establishing a deep recurrent model to predict and eliminate the PHT component of hand motion. More specifically, we propose a machine learning-based, assumption-free, and real-time PHT elimination framework, the PHTNet, by incorporating deep bidirectional recurrent neural networks. The PHTNet is developed over a hand motion dataset of 81 ET and PD patients collected systematically in a movement disorders clinic over 3 years. The PHTNet is the first intelligent systems model developed on this scale for PHT elimination that maximizes the resolution of estimation and allows for prediction of future and upcoming sub-movements.

Overview

CoChat Summary

TL;DR

PHTNet is a deep bidirectional recurrent neural network for estimating and predicting the voluntary component of pathological hand tremor, trained on a large single-centre dataset of 81 PD/ET patients. The paper reports that PHTNet outperforms BMFLC, EBMFLC, and WAKE on pseudo-synthesized validation/test data and shows generalization on real action tremor and healthy-control recordings.

Key findings

  • The dataset used to develop PHTNet included 81 PD and ET individuals from a single-centre movement disorders clinic; the paper also reports 87.5 hours of data from 81 patients and 250 tremor assessments in total.
  • For model development, recordings were split by subject into 49 training, 16 validation, and 16 test subjects; data were downsampled to 100 Hz.
  • On pseudo-synthesized validation and test sets, PHTNet was evaluated over 24,300 samples per set and 48,600 input-output pairs in total for predictive-behavior analysis; the paper states that PHTNet showed significantly better performance than BMFLC, EBMFLC, and WAKE using ANOVA and pairwise Z-tests.
  • The authors report a 4-layer GRU/BiRNN architecture, 400 features as best for hidden representation, and 4-second input sequences as the best accuracy/computation tradeoff; the model uses MSE training with Adam at 0.0001 learning rate.
  • In the healthy-control experiment, the paper reports recordings from 2 healthy male volunteers aged 28 and 37, and says PHTNet’s estimated signals “perfectly match” the inputs.

Methods

  • The paper proposes PHTNet, a deep bidirectional recurrent neural network built from GRU cells in a sequence-to-sequence setup, with separate forward (online/predictive) and backward (offline) processing pipelines and no mixing matrix.
  • Training and evaluation used a subject-level split of clinical hand-motion recordings from 81 PD/ET patients collected at the London Movement Disorders Centre; tasks included 7 scripted 20-second tasks over three trials, with static tremor from 6 tasks and action tremor from a finger-to-nose task.
  • To obtain ground truth for action tremor, the authors used an inverse evaluation method: they created pseudo-synthesized action tremor by mixing synthesized voluntary sinusoidal components with real static tremor recordings; performance was quantified with MSE, and statistical comparisons used ANOVA, pairwise Z-tests, and D’Agostino–Pearson normality tests.

Limitations

  • The paper states that the achieved horizon of prediction was limited and that it is working on improving prediction by augmenting input space and expanding the dataset.
  • The authors note that the deep model requires relatively stronger computational support, which may limit deployment when computational resources are constrained.
  • Evident from the text: the clinical dataset is from a single centre and the healthy-control validation used only 2 volunteers; these factors may limit external generalizability (evident from the text).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41598-020-58912-9

Cites

74 of 74 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 (a) Illustration of the 7 scripted tasks performed by PD and ET patients for each tremor assessment. Please note that, these are representative pictures (not including any patients). 1) Rest-1; 2) Rest-2; 3) Posture-1; 4) Posture-2; 5) Action tremor (repetitive finger to nose motion); 6) Load-1 (empty cup); 7) Load-2 (1-lb weight in the cup). (b) Placement of the 3-axis accelerometer sensor on the dorsum of a hand.
Figure 1 (a) Illustration of the 7 scripted tasks performed by PD and ET patients for each tremor assessment. Please note that, these are representative pictures (not including any patients). 1) Rest-1; 2) Rest-2; 3) Posture-1; 4) Posture-2; 5) Action tremor (repetitive finger to nose motion); 6) Load-1 (empty cup); 7) Load-2 (1-lb weight in the cup). (b) Placement of the 3-axis accelerometer sensor on the dorsum of a hand.CC BY · PubMed Central
Figure 2 (a) A schematic of a one-layer RNN within the PHTNet with the unfolded version demonstrated on the right-hand side, which clearly shows the processing pipeline for different time instances. For the schematic on the left-hand side, it should be noted that the branch denoted by the weight W also applies one sample delay in time. (b) A gated recurrent unit (GRU) which is employed in PHTNet as the recurrent cell and is equipped with reset gate (r) and update gate (z). (c) The architecture of PHTNet, which is a 4-layer deep bidirectional recurrent neural network. h← defines the backward cells for offline tremor elimination, and h→ defines the forward cell of the network for online tremor estimation/prediction. As shown in the diagram, the forward path is completely distinct from the backward path and their outputs are not merged into a single output sequence. The red gradient in the output blocks represents the degree of error in the extracted voluntary component. The high intensity of the red color symbolizes a high degree of error and the opposite mimics lower error rates. (d) The overall workflow of the proposed framework. Note that the voluntary component is recalculated for the next time instance and then is compared with the output of the network. This strategy is taken to enable the network with predictive features.
Figure 2 (a) A schematic of a one-layer RNN within the PHTNet with the unfolded version demonstrated on the right-hand side, which clearly shows the processing pipeline for different time instances. For the schematic on the left-hand side, it should be noted that the branch denoted by the weight W also applies one sample delay in time. (b) A gated recurrent unit (GRU) which is employed in PHTNet as the recurrent cell and is equipped with reset gate (r) and update gate (z). (c) The architecture of PHTNet, which is a 4-layer deep bidirectional recurrent neural network. h← defines the backward cells for offline tremor elimination, and h→ defines the forward cell of the network for online tremor estimation/prediction. As shown in the diagram, the forward path is completely distinct from the backward path and their outputs are not merged into a single output sequence. The red gradient in the output blocks represents the degree of error in the extracted voluntary component. The high intensity of the red color symbolizes a high degree of error and the opposite mimics lower error rates. (d) The overall workflow of the proposed framework. Note that the voluntary component is recalculated for the next time instance and then is compared with the output of the network. This strategy is taken to enable the network with predictive features.CC BY · PubMed Central
Figure 3 Comparison of normalized MSE (y-axis) over validation and test sets for different number of hidden layers (x-axis). Performance of the PHTNet in each test case is shown in boxplots, where the orange line indicates the median, the box indicates the range between 25% to 75% quartiles, and the lines indicate the standard deviation range. “Prediction Error” indicates the error for only the last sample of estimated voluntary signal. “Forward” and “Backward” indicate which path of the PHTNet is used.
Figure 3 Comparison of normalized MSE (y-axis) over validation and test sets for different number of hidden layers (x-axis). Performance of the PHTNet in each test case is shown in boxplots, where the orange line indicates the median, the box indicates the range between 25% to 75% quartiles, and the lines indicate the standard deviation range. “Prediction Error” indicates the error for only the last sample of estimated voluntary signal. “Forward” and “Backward” indicate which path of the PHTNet is used.CC BY · PubMed Central
Figure 4 Comparison of Normalized Mean Squared Error (y-axis) over validation and test sets for different lengths of input sequence in seconds (x-axis). Performance of the PHTNet in each test case is shown in boxplots, where the orange line indicates the median, the box indicates the range between 25% to 75% quartiles, and the lines indicate the standard deviation range. “Prediction Error” indicates the error for only the last sample of estimated voluntary signal. “Forward” and “Backward” indicate which path of the PHTNet is used.
Figure 4 Comparison of Normalized Mean Squared Error (y-axis) over validation and test sets for different lengths of input sequence in seconds (x-axis). Performance of the PHTNet in each test case is shown in boxplots, where the orange line indicates the median, the box indicates the range between 25% to 75% quartiles, and the lines indicate the standard deviation range. “Prediction Error” indicates the error for only the last sample of estimated voluntary signal. “Forward” and “Backward” indicate which path of the PHTNet is used.CC BY · PubMed Central
Algorithm 1 The algorithmic overview of the development phase of the PHTNet.
Algorithm 1 The algorithmic overview of the development phase of the PHTNet.CC BY · PubMed Central
Figure 5 Visualization of the network’s output when the pseudo-synthesized evaluation signals from validation and test sets are fed to the network. Please note that due to the pseudo-synthesized nature of these signals, the known synthetic ground truth is mixed with pathological data to augment the input space and enforce the model to learn how to extract the pathological tremorous motion. The PSD of input and output signals are also included to demonstrate the transformation and manipulation of the spectral contents. The PSD plots are obtained by sliding a Fast Fourier Transform (FFT) window of size 50 samples and overlap size of 45 samples over the signals. The color bars also determine the density of each frequency at each time and its dimension is Amplitude²/Hz. (a,b) Two instances of the synthesized evaluation signals from the patients in the validation set. (c,d) Two instances of the evaluation signals generated from the static tremor recordings of patients in the test set. Please also note that the presented PSD diagrams are only obtained for visualization purposes and their properties do not influence any parameter or hyper-parameter of the PHTNet.
Figure 5 Visualization of the network’s output when the pseudo-synthesized evaluation signals from validation and test sets are fed to the network. Please note that due to the pseudo-synthesized nature of these signals, the known synthetic ground truth is mixed with pathological data to augment the input space and enforce the model to learn how to extract the pathological tremorous motion. The PSD of input and output signals are also included to demonstrate the transformation and manipulation of the spectral contents. The PSD plots are obtained by sliding a Fast Fourier Transform (FFT) window of size 50 samples and overlap size of 45 samples over the signals. The color bars also determine the density of each frequency at each time and its dimension is Amplitude²/Hz. (a,b) Two instances of the synthesized evaluation signals from the patients in the validation set. (c,d) Two instances of the evaluation signals generated from the static tremor recordings of patients in the test set. Please also note that the presented PSD diagrams are only obtained for visualization purposes and their properties do not influence any parameter or hyper-parameter of the PHTNet.CC BY · PubMed Central
Figure 6 (a) Comparison of the performance between BMFLC, EBMFLC, WAKE, and PHTNet on the validation and test sets. The red line in the boxplots indicates the median performance and the box indicates 25%% and 75% quartiles. The dashed lines show standard deviation and the red crosses show outliers. (b,c) Multiple comparison of the mean performance of the four methods over the validation and test sets. The dashed lines indicate the upper and lower 95% confidence bounds for the estimation error of each method and the y-axis represents the labels of compared PHT removal techniques.
Figure 6 (a) Comparison of the performance between BMFLC, EBMFLC, WAKE, and PHTNet on the validation and test sets. The red line in the boxplots indicates the median performance and the box indicates 25%% and 75% quartiles. The dashed lines show standard deviation and the red crosses show outliers. (b,c) Multiple comparison of the mean performance of the four methods over the validation and test sets. The dashed lines indicate the upper and lower 95% confidence bounds for the estimation error of each method and the y-axis represents the labels of compared PHT removal techniques.CC BY · PubMed Central
Figure 7 (a,b) Visualization of the 95% confidence boundaries for the spectral contents of input-output pairs in validation and test sets, respectively. The solid lines indicate the mean of spectral contents and the highlighted area indicate confidence boundaries. Also note that the y-axis on the right represents the p−value between the spectral contents of the input and output signals. (c,d) Visualization of standard deviation boundaries for input-output pairs along with the mean of spectral contents of input and output signals in validation and test sets, respectively. Please note that the spectral contents of the input signals are shown in “red” color, while those of the output signals are depicted with “blue” color. The dominant activity at around 5 Hz is only observed in the input signals (“red” color), which is expected due to the strong power of pathological tremor around 5 Hz.
Figure 7 (a,b) Visualization of the 95% confidence boundaries for the spectral contents of input-output pairs in validation and test sets, respectively. The solid lines indicate the mean of spectral contents and the highlighted area indicate confidence boundaries. Also note that the y-axis on the right represents the p−value between the spectral contents of the input and output signals. (c,d) Visualization of standard deviation boundaries for input-output pairs along with the mean of spectral contents of input and output signals in validation and test sets, respectively. Please note that the spectral contents of the input signals are shown in “red” color, while those of the output signals are depicted with “blue” color. The dominant activity at around 5 Hz is only observed in the input signals (“red” color), which is expected due to the strong power of pathological tremor around 5 Hz.CC BY · PubMed Central
Figure 8 Visualization of the predictive capability of the network over special test signals.
Figure 8 Visualization of the predictive capability of the network over special test signals.CC BY · PubMed Central
Figure 9 Visualization of the network output when real action tremor signals are fed to the network. Our method was compared with three other methods, referred to as BMFLC, EBMFLC, and WAKE. The details of each patient whose signal is shown here are as follows. (a) [ET - Right hand - de novo - Training]. (b) [PD - Right hand - Under treatment - Validation]. (c) [PD - Left hand - de novo - Test]. (d) [ET - Right hand - Under treatment - Test]. (e) [PD - Right hand - Under treatment - Test]. (f) [PD - Right hand - de novo - Test]. (g) [PD - Right hand - Under treatment - Test]. (h) [ET - Right hand - de novo - Training]. (i) [PD - Right hand - Under treatment - Training]. (j) [PD - Right hand - de novo - Training]. (k) [ET - Left hand - Under treatment - Validation]. (l) [PD - Left hand - Under treatment - Validation].
Figure 9 Visualization of the network output when real action tremor signals are fed to the network. Our method was compared with three other methods, referred to as BMFLC, EBMFLC, and WAKE. The details of each patient whose signal is shown here are as follows. (a) [ET - Right hand - de novo - Training]. (b) [PD - Right hand - Under treatment - Validation]. (c) [PD - Left hand - de novo - Test]. (d) [ET - Right hand - Under treatment - Test]. (e) [PD - Right hand - Under treatment - Test]. (f) [PD - Right hand - de novo - Test]. (g) [PD - Right hand - Under treatment - Test]. (h) [ET - Right hand - de novo - Training]. (i) [PD - Right hand - Under treatment - Training]. (j) [PD - Right hand - de novo - Training]. (k) [ET - Left hand - Under treatment - Validation]. (l) [PD - Left hand - Under treatment - Validation].CC BY · PubMed Central
Figure 10 Visualization of the network output when real action tremor signals are fed to the network. In this case, the prediction samples are employed. The details of each patient whose signal is shown here are as follows. (a) [ET - Right hand - de novo - Training]. (b) [PD - Right hand - Under treatment - Validation]. (c) [PD - Left hand - de novo - Test]. (d) [ET - Right hand - Under treatment - Test]. (e) [PD - Right hand - Under treatment - Test]. (f) [PD - Right hand - de novo - Test]. (g) [PD - Right hand - Under treatment - Test]. (h) [ET - Right hand - de novo - Training]. (i) [PD - Right hand - Under treatment - Training]. (j) [PD - Right hand - de novo - Training]. (k) [ET - Left hand - Under treatment - Validation]. (l) [PD - Left hand - Under treatment - Validation].
Figure 10 Visualization of the network output when real action tremor signals are fed to the network. In this case, the prediction samples are employed. The details of each patient whose signal is shown here are as follows. (a) [ET - Right hand - de novo - Training]. (b) [PD - Right hand - Under treatment - Validation]. (c) [PD - Left hand - de novo - Test]. (d) [ET - Right hand - Under treatment - Test]. (e) [PD - Right hand - Under treatment - Test]. (f) [PD - Right hand - de novo - Test]. (g) [PD - Right hand - Under treatment - Test]. (h) [ET - Right hand - de novo - Training]. (i) [PD - Right hand - Under treatment - Training]. (j) [PD - Right hand - de novo - Training]. (k) [ET - Left hand - Under treatment - Validation]. (l) [PD - Left hand - Under treatment - Validation].CC BY · PubMed Central
Figure 11 (a,b) Results of applying PHTNet on the recordings of static posture tasks and finger-to-nose motion test for the two healthy subjects.
Figure 11 (a,b) Results of applying PHTNet on the recordings of static posture tasks and finger-to-nose motion test for the two healthy subjects.CC BY · PubMed Central

Verification

Claims the paper makes

5 Valid · 5 Not valid · 3 Not checked · 56 Own results and interpretations

18 more claims

Create a free account to see all 69 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 31
  • Probably not a substantial claim. 25
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 74 references

Research works printed but not linked: 1

  1. Goodfellow, I., Courville, A. & Bengio, Y. Deep learning, vol. 1 (MIT press Cambridge, 2016).

Cited by

See all 33

Metadata

Authors
Shahtalebi, S., Atashzar, S. F., Samotus, O., Patel, R. V., Jog, M. S., & Mohammadi, A.
Affiliations
  • Concordia University (CA)
  • New York University (US)
  • London Health Sciences Centre (CA)
  • Western University (CA)
Corresponding author
Mohammadi, Arash
Venue
Scientific Reports
Published
2020-02-10
Volume
10
Issue
1
Pages
2195-2195
ISSN
2045-2322 (ISSN-L)
OpenAlex
W3004983582
Cited by
32
References
74
Citations by year
2026: 2 · 2025: 8 · 2024: 1 · 2023: 6 · 2022: 8 · 2021: 4 · 2020: 3
FWCI
2.82
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
Gouvernement du Canada | Natural Sciences and Engineering Research Council of Canada (RGPIN-2016-049988)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Motor Control and Adaptation
  • Health Sciences › Medicine › Clinical Neurology › Deep Brain Stimulation and Movement Disorders
  • Health Sciences › Health Professions › Speech and Hearing › Voice and Speech Disorders
Keywords
Parkinson's disease, essential tremor, bidirectional recurrent neural network, hand movement, movement disorders, rehabilitation technology, assistive technology
MeSH terms
Machine Learning, Aged, Aged, 80 and over, Female, Hand, Humans, Male, Middle Aged, Motion, Movement, Parkinson Disease, Prognosis, Quality of Life, Tremor, Neural Networks, Computer, Essential Tremor

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: PHTNet: Characterization and Deep Mining of Involuntary Pathological Hand Tremor using Recurrent Neural Network Models. https://app.cochat.ai/papers/10.1038/s41598-020-58912-9
The paper
Shahtalebi, S., Atashzar, S. F., Samotus, O., Patel, R. V., Jog, M. S., & Mohammadi, A. (2020). PHTNet: Characterization and Deep Mining of Involuntary Pathological Hand Tremor using Recurrent Neural Network Models. Scientific Reports. https://doi.org/10.1038/s41598-020-58912-9

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.