Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IdentiARAT: Toward Automated Identification of Individual ARAT Items from Wearable Sensors

Daniel Homm, Patrick Carqueville, Christian Eichhorn, Thomas Weikert, Thomas Menard, David A. Plecher, Chris Awai Easthope

TL;DR

The paper addresses automating the labeling of ARAT items using wrist-worn IMUs to reduce clinician workload and subjectivity in upper-limb assessment. It adopts MiniROCKET, a fast time-series classifier, applied to $60\,\text{Hz}$ IMU data from two wrist sensors, and evaluates performance with $10$-fold cross-validation, reporting domain-level accuracy of $82.3\%$ and item-level accuracy around $61\%$ after removing the longest sequences, with domain accuracy reaching $93.9\%$. Key contributions include a grid-search over preprocessing, demonstration that multi-stream sensor inputs improve accuracy (best ~$47.2\%$ for item-level on the left wrist) and that longer sequences degrade performance, highlighting the trade-off between data richness and temporal variability. The findings show feasibility for home-use auto-labeling and clinical decision support, while also outlining clear directions for improvements, such as incorporating additional sensors or transformer-based models to achieve finer-grained item discrimination.

Abstract

This study explores the potential of using wrist-worn inertial sensors to automate the labeling of ARAT (Action Research Arm Test) items. While the ARAT is commonly used to assess upper limb motor function, its limitations include subjectivity and time consumption of clinical staff. By using IMU (Inertial Measurement Unit) sensors and MiniROCKET as a time series classification technique, this investigation aims to classify ARAT items based on sensor recordings. We test common preprocessing strategies to efficiently leverage included information in the data. Afterward, we use the best preprocessing to improve the classification. The dataset includes recordings of 45 participants performing various ARAT items. Results show that MiniROCKET offers a fast and reliable approach for classifying ARAT domains, although challenges remain in distinguishing between individual resembling items. Future work may involve improving classification through more advanced machine-learning models and data enhancements.

IdentiARAT: Toward Automated Identification of Individual ARAT Items from Wearable Sensors

TL;DR

The paper addresses automating the labeling of ARAT items using wrist-worn IMUs to reduce clinician workload and subjectivity in upper-limb assessment. It adopts MiniROCKET, a fast time-series classifier, applied to IMU data from two wrist sensors, and evaluates performance with -fold cross-validation, reporting domain-level accuracy of and item-level accuracy around after removing the longest sequences, with domain accuracy reaching . Key contributions include a grid-search over preprocessing, demonstration that multi-stream sensor inputs improve accuracy (best ~ for item-level on the left wrist) and that longer sequences degrade performance, highlighting the trade-off between data richness and temporal variability. The findings show feasibility for home-use auto-labeling and clinical decision support, while also outlining clear directions for improvements, such as incorporating additional sensors or transformer-based models to achieve finer-grained item discrimination.

Abstract

This study explores the potential of using wrist-worn inertial sensors to automate the labeling of ARAT (Action Research Arm Test) items. While the ARAT is commonly used to assess upper limb motor function, its limitations include subjectivity and time consumption of clinical staff. By using IMU (Inertial Measurement Unit) sensors and MiniROCKET as a time series classification technique, this investigation aims to classify ARAT items based on sensor recordings. We test common preprocessing strategies to efficiently leverage included information in the data. Afterward, we use the best preprocessing to improve the classification. The dataset includes recordings of 45 participants performing various ARAT items. Results show that MiniROCKET offers a fast and reliable approach for classifying ARAT domains, although challenges remain in distinguishing between individual resembling items. Future work may involve improving classification through more advanced machine-learning models and data enhancements.

Paper Structure

This paper contains 22 sections, 1 equation, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Study design
  4. Data
  5. Analytics
  6. Experiments
  7. Validation
  8. Data Preprocessing
  9. Input Features
  10. Domain-Level Classification
  11. Sequence Length Impact
  12. Results
  13. Data Preprocessing
  14. Input Features
  15. Domain-Level Classification
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Patient performing ARAT item, Cube 10cm
  • Figure 2: RandOm Convolutional KErnel Transform (ROCKET) approach Rocket-Figure.
  • Figure 3: Confusion matrix of the aggregated validation results for the all-item classification, using the complete left wrist
  • Figure 4: Confusion matrix of the aggregated validation results for the item-domain classification, using the complete left wrist data.