CoChat

The DOI matches its Crossref record; 20 of the 20 references it cites were found in a registry or an index; 6 of the 13 sources its claims cite were checked for support.

Explainable artificial intelligence for searching frequency characteristics in Parkinson's disease tremor

Open access
Lee, Rui En, Chan, Ping Yi · 2023 · Scientific Reports

Abstract

The distinction between Parkinson's disease (PD) and essential tremor (ET) tremors is subtle, posing challenges in differentiation. To accurately classify the PD and ET, BiLSTM-based recurrent neural networks are employed to classify between normal patients (N), PD patients, and ET patients using accelerometry data on their lower arm (L), hand (H), and upper arm (U) as inputs. The trained recurrent neural network (RNN) has reached 80% accuracy. The neural network is analyzed using layer-wise relevance propagation (LRP) to understand the internal workings of the neural network. A novel explainable AI method, called LRP-based approximate linear weights (ALW), is introduced to identify the similarities in relevance when assigning the class scores in the neural network. The ALW functions as a 2D kernel that linearly transforms the input data directly into the class scores, which significantly reduces the complexity of analyzing the neural network. This new classification method reconstructs the neural network's original function, achieving a 73% PD and ET tremor classification accuracy. By analyzing the ALWs, the correlation between each input and the class can also be determined. Then, the differentiating features can be subsequently identified. Since the input is preprocessed using short-time Fourier transform (STFT), the differences between the magnitude of tremor frequencies ranging from 3 to 30 Hz in the mean N, PD, and ET subjects are successfully identified. Aside from matching the current medical knowledge on frequency content in the tremors, the differentiating features also provide insights about frequency contents in the tremors in other frequency bands and body parts.

Overview

CoChat Summary

TL;DR

This paper trains a BiLSTM-based recurrent neural network to classify normal, Parkinson’s disease, and essential tremor subjects from upper-limb accelerometry, then applies layer-wise relevance propagation and a new approximate linear weights method to identify frequency-domain differentiating features. The RNN reached 80% accuracy, while the simplified ALW classifier reached 73% accuracy and highlighted tremor frequency bands across lower arm, hand, and upper arm.

Key findings

  • The BiLSTM-based RNN classified N, PD, and ET with 80% accuracy; the LRP-based ALW reconstruction achieved 73% accuracy.
  • LRP indicated that only the first and last 20 timesteps were important for classification: accuracy dropped when fewer than 20 preserved timesteps were used and stayed relatively constant with more than 20.
  • ALW identified frequency features including PD with high amplitude at 4.7 Hz in the lower arm and 4.7 Hz and 5.5 Hz in the hand and upper arm; ET with high amplitude at 4.7–5.5 Hz, 7.8–8.6 Hz, and 10.9–11.7 Hz in the lower arm, 4.7–8.6 Hz in the hand, and 6.3–9.4 Hz and 10.9–11.3 Hz in the upper arm; and N with relatively higher amplitude at 20–30 Hz at all body parts and at 3.1 Hz.
  • Erasure analysis supported the LRP implementation: removing positively relevant parameters dropped accuracy to almost zero, while removing negatively relevant parameters raised accuracy to 100%.

Methods

  • Tremors were measured with an SBG IG-500A attitude and heading reference system on the lower arm, hand, and upper arm during REST, OUT, WING, DRINK, and FNF actions.
  • The methodology reports 67 N, 87 PD, and 18 ET subjects; a later comparison section lists 86 PD patients (55 males, 31 females, mean age 69.67 ± 8.18), 18 ET patients (10 males, eight females, mean age 67.06 ± 8.93), and 67 normal subjects (35 males, 32 females, mean age 55.03 ± 10.8). Data were interpolated with PCHIP, bandpass filtered at 3–30 Hz, and preprocessed with STFT at 4.7 ms time resolution and 0.7825 Hz frequency resolution from 3.12 to 29.69 Hz.
  • A BiLSTM-based RNN was trained with a class-balanced testing set of 30, an 80:20 training-validation split for the remaining data, batch normalization, and 40% dropout. LRP was applied to explain the network, and ALW approximated the input-to-class-score relationship as a 2D convolutional layer.

Limitations

  • The paper states that ALW neglects neural-network nonlinearity and lowers accuracy; LRP has relevance leakage from bias terms and the LSTM forget gate; limited datasets and oversampling of ET data may have caused overfitting; STFT captures stationary and linear data although tremor is described as non-stationary and non-linear; drinking action may introduce dynamic action tremor noise; and the reported unique features need further validation with clinicians’ judgment.
  • Evident from the text: the ET group is much smaller than the N and PD groups, and the normal group has a lower mean age than the PD and ET groups, which may affect generalizability.

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

Integrity signals

Registry

Crossref · 10.1038/s41598-023-45802-z

Cites

20 of 20 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Methodology flowchart.
Figure 1 Methodology flowchart.CC BY · PubMed Central
Figure 2 BiLSTM-based neural network architecture.
Figure 2 BiLSTM-based neural network architecture.CC BY · PubMed Central
Figure 3 Visualization of ALW classification operation.
Figure 3 Visualization of ALW classification operation.CC BY · PubMed Central
Figure 4 Visualisation of ALW classification operation.
Figure 4 Visualisation of ALW classification operation.CC BY · PubMed Central
Figure 5 Examples of splitting the approximate weights into quarters. (i) Approximate weights of input parameter 2 for class N, which corresponds to 3.91 Hz at the lower arm. The quartered signs are NPNN. (ii) Approximate weights of input parameter 2 for class PD, which corresponds to 3.91 Hz at the lower arm. The quartered signs are PNPP.
Figure 5 Examples of splitting the approximate weights into quarters. (i) Approximate weights of input parameter 2 for class N, which corresponds to 3.91 Hz at the lower arm. The quartered signs are NPNN. (ii) Approximate weights of input parameter 2 for class PD, which corresponds to 3.91 Hz at the lower arm. The quartered signs are PNPP.CC BY · PubMed Central
Figure 6 Plots of relevance against timesteps for each sample subject from class 1 (N), class 2 (PD), and class 3 (ET). Different lines refer to the relevance of each of the 105 sets of input parameters.
Figure 6 Plots of relevance against timesteps for each sample subject from class 1 (N), class 2 (PD), and class 3 (ET). Different lines refer to the relevance of each of the 105 sets of input parameters.CC BY · PubMed Central
Figure 7 Graph of accuracy vs number of start and end timesteps preserved.
Figure 7 Graph of accuracy vs number of start and end timesteps preserved.CC BY · PubMed Central
Figure 8 Heatmaps of relevance from class 1 (N), class 2 (PD), and class 3 (ET). The x-axis corresponds to the timesteps, and the y-axis corresponds to the frequency in top-down ascending order from 3.1 Hz to 28.9 Hz, stacked for the lower arm, hand, and upper arm. Only the first and last 20 timesteps are plotted since the middle timesteps have relevances of close to 0.
Figure 8 Heatmaps of relevance from class 1 (N), class 2 (PD), and class 3 (ET). The x-axis corresponds to the timesteps, and the y-axis corresponds to the frequency in top-down ascending order from 3.1 Hz to 28.9 Hz, stacked for the lower arm, hand, and upper arm. Only the first and last 20 timesteps are plotted since the middle timesteps have relevances of close to 0.CC BY · PubMed Central
Figure 9 Graph of accuracies against the proportion of input parameters zeroed.
Figure 9 Graph of accuracies against the proportion of input parameters zeroed.CC BY · PubMed Central
Figure 10 Bar chart of weights summed across all timesteps of each frequency for lower arm, hand and upper arm. The sum of weights are normalised to range (− 1, 1) across classes for better comparison.
Figure 10 Bar chart of weights summed across all timesteps of each frequency for lower arm, hand and upper arm. The sum of weights are normalised to range (− 1, 1) across classes for better comparison.CC BY · PubMed Central
Figure 11 Box plot of average tremor magnitude of frequency 5.47 Hz.
Figure 11 Box plot of average tremor magnitude of frequency 5.47 Hz.CC BY · PubMed Central

Verification

Claims the paper makes

3 Valid · 2 Not valid · 7 Not checked · 37 Own results and interpretations

14 more claims

Create a free account to see all 49 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 21
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 20 references

Cited by

See all 13

Metadata

Authors
Lee, R. E., & Chan, P. Y.
Affiliations
Monash University Malaysia (MY)
Corresponding author
Chan, Ping Yi
Venue
Scientific Reports
Published
2023-10-30
Volume
13
Issue
1
Pages
18622-18622
ISSN
2045-2322 (ISSN-L)
OpenAlex
W4388023898
Cited by
13
References
18
Citations by year
2026: 4 · 2025: 4 · 2024: 5
FWCI
1.22
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Monash University Research Portal (Monash University) (submittedVersion, other-oa)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Health Sciences › Medicine › Clinical Neurology › Deep Brain Stimulation and Movement Disorders
Keywords
Parkinson's disease, essential tremor, tremor classification, explainable AI, layer-wise relevance propagation, BiLSTM, accelerometry, short-time Fourier transform, frequency bands, differential diagnosis
MeSH terms
Artificial Intelligence, Birth Weight, Humans, Parkinson Disease, 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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (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: Explainable artificial intelligence for searching frequency characteristics in Parkinson's disease tremor. https://app.cochat.ai/papers/10.1038/s41598-023-45802-z
The paper
Lee, R. E., & Chan, P. Y. (2023). Explainable artificial intelligence for searching frequency characteristics in Parkinson's disease tremor. Scientific Reports. https://doi.org/10.1038/s41598-023-45802-z

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.