Currently submitted to: JMIR mHealth and uHealth
Date Submitted: Jun 7, 2026
Open Peer Review Period: Jun 8, 2026 - Aug 3, 2026
(currently open for review)
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Cross-validation of wearable device-based machine-learning algorithms for the estimation of energy expenditure during semi-structured wheelchair activities: a pilot study
ABSTRACT
Background:
Most energy expenditure (EE) estimation algorithms for wheelchair users (WCUs) have been developed for individuals with spinal cord injuries and rely on data from structured laboratory settings.
Objective:
(1) To evaluate the accuracy of wearable-based machine-learning EE estimation algorithms in a heterogeneous group of WCUs during semi-structured wheelchair activities. (2) To examine the impact of time segmentation and sensor location on model performance.
Methods:
Pilot data from nine WCUs with different disabilities (age: 45.8 ± 14.9 years, body-mass: 67.9 ± 19.9 kg) was collected during seated rest and a semi-structured activity course simulating daily life wheelchair activities (semi-structured dataset). Input data for the algorithms was provided by a chest-strap HR monitor, two inertial measurement units (non-dominant wrist & wheel), and information on personal characteristics (sex, age, body mass, height, physical activity level). We compared the EE estimation performance of seven machine-learning models (two linear, five nonlinear: Support Vector Regression (SVR), Random Forest Regression, eXtreme Gradient Boosting, Neural Networks and Gaussian Process Regression (GPR)), trained on three datasets: structured lab data (collected previously, data of 20 WCUs), semi-structured data, and a combined dataset. Models were tested on the semi-structured dataset and performance evaluated with the mean absolute percentage error (MAPE) and coefficient of determination (R²).
Results:
Models trained on structured lab data and tested on semi-structured data showed poor to moderate accuracy (MAPE > 20%, R² < 0.7 for all models). Training and testing on semi-structured data using leave-one-participant-out cross-validation improved performance (MAPE < 20%, R² > 0.7), though none achieved MAPE < 10%, a benchmark for acceptable accuracy. Combining datasets did not enhance model performance. Among algorithms, GPR and SVR were most stable. Including features from the wrist-mounted IMU outperformed wheel-only and combined sensor setups. Longer segment lengths yielded slightly more stable estimates, but 30-sec segments better captured dynamic EE fluctuations.
Conclusions:
Generalizing EE models across structured and semi-structured settings remains challenging. Training on semi-structured data with wrist-mounted IMU sensors provided the most accurate results. While longer segments offered stability, 30-sec segments are recommended for capturing transient EE changes. Future work should explore personalized modeling using larger datasets and subgroup-specific models based on demographic, impairment- and fitness characteristics. Clinical Trial: NA
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.