Cross-Dataset Evaluation of Sit-to-Stand Movement Classifiers for Post-Stroke Rehabilitation

What will it take to get thru your thick skulls that 'assessments' do nothing for recovery unless THEY POINT DIRECTLY TO EXACT RECOVERY PROTOCOLS?

This did nothing towards that, so useless!

• sit-to-stand (25 posts to December 2012)
  

Once again more research on sit-to-stand and NOTHING that gets survivors recovered! Hope you like being disabled when you become the 1 in 4 per WHO that has a stroke! 

 Cross-Dataset Evaluation of Sit-to-Stand Movement Classifiers for Post-Stroke Rehabilitation

  L. Palumbi1*, T. Kuhlgatz1, M. Caruso1,2, T. Seel1 and L. Budde1 1Leibniz University Hannover, Institute of Mechatronic Systems, Hanover, Germany 2Center for Automation and Robotics, Spanish National Research Council (CSIC-UPM), Madrid, Spain *Corresponding author, email: lucia.palumbi@imes.uni-hannover.de 

 Abstract: 

Automated assessment could support at-home post-stroke rehabilitation, yet ensuring cross-dataset generalizability is critical for real-world adoption. This study evaluates rule-based and Random Forest classifiers, trained on XSense IMU data, against the independent CeTI-Locomotion dataset. Zero-shot evaluation demonstrated the robustness of the rule-based approach (71.3% accuracy) compared to Random Forest (17.8%), which significantly improved to 58.2% with one-shot calibration. These findings indicate that generalizability is achievable through biomechanically grounded or adaptive strategies, marking a key step toward robust, clinically deployable rehabilitation systems. © 2026 Lucia Palumbi; licensee Infinite Science Publishing This is an Open Access article distributed under the terms of the Creative Commons Attribution License CC-BY 4.0., which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 

 I. Introduction Stroke remains the third leading cause of disability worldwide [1]. To regain mobility and functional independence, post-stroke patients require high-dose, repetitive rehabilitation exercises. However, limited access to supervised physiotherapy and patient non-compliance between sessions significantly hinder recovery outcomes [2]. Wearable sensor-based systems for automated movement assessment offer a promising solution by enabling unsupervised monitoring during home-based rehabilitation. Recent studies have demonstrated the feasibility of using inertial measurement units (IMUs) to assess rehabilitation exercises. E.g., Komaris et al. [3] showed that IMU-based metrics movement smoothness, intensity, and consistency – could effectively evaluate exercise performance at-home, while Ranganathan et al. [4] achieved 90% accuracy in detecting compensatory upper-limb movements during reaching and manipulation tasks. In our earlier work [5], we demonstrated the feasibility of distinguishing sit-to-stand (STS) movements from characteristic related post-stroke impairments, including asymmetrical weight-bearing, lateral trunk deviation, and compensatory stepping [6]. Using 6D IMUs, we evaluated both rule-based and learning-based classification models, which achieved average accuracies of 89.78% and 94.03%, respectively. However, these approaches share a critical limitation: validation is performed exclusively on the same datasets used for model development, leaving cross-dataset generalizability untested. Algorithms trained and tested solely on small, controlled datasets often degrade significantly when applied to the heterogeneous movement patterns of real-life settings [7]. To eliminate the need for subject-specific calibration, reliable monitoring systems must be robust against these shifts in distribution. In this work, we evaluate the cross-dataset transferability of our STS classifiers when applied out-of-the-box to the CeTI-Locomotion dataset [8], which contains IMU data from 50 healthy adults performing sit-to-stand. We assess both zero-shot and one-shot adaptation approaches, quantify the domain gap between training and target distributions and test the model’s ability to distinguish natural kinematic variability from pathology, which is essential to ensure specificity and minimize false positives in clinical practice [2]. II. Material and methods Datasets. We evaluated classifiers trained on the IMU dataset collected in [5] against the CeTI-Locomotion dataset [8]. The source domain data comprised 330 STS trials from 11 healthy participants mimicking post-stroke patterns under physiotherapist supervision: correct execution, asymmetrical movement simulating hemiparesis (plegie), uncontrolled descent or failure of completing the movement (fail), and compensatory stepping (step). The target dataset, CeTI-Locomotion, contains 499 STS repetitions from 50 healthy adults recorded with the IMU based Rokoko Smartsuit Pro motion capture system, at 100 Hz. Since CeTI-Locomotion contains only correct executions, if our algorithms [5] generalize well, they should classify all trials in the CeTI_Locomotion dataset as correct, making misclassifications direct measures of false positive compensation movements.