Related Experiment Video
Updated: Sep 18, 2026
Methodology for Establishing a Community-Wide Life Laboratory for Capturing Unobtrusive and Continuous Remote Activity and Health Data
Published on: July 27, 2018
Classification of daily activities using a wireless instrumented insole (WalkinSense) in a semi free-living setting
Anne Backes1, Melanie Eckelt1,2, Jennifer Fayad1,2
1Department of Precision Health, Luxembourg Institute of Health, Strassen, Luxembourg.
Background:
Accurate monitoring of activities of daily living (ADLs) in real‑world environments is essential for preventive care and rehabilitation, yet it remains difficult to achieve outside controlled laboratory settings. Instrumented insoles provide a promising, unobtrusive solution for continuous monitoring. However, the use of multimodal systems integrating plantar pressure and inertial data in naturalistic conditions and larger cohorts is still limited. This study therefore aims to evaluate the accuracy of a wireless instrumented insole (WalkinSense) that fuses pressure and inertial sensor data to classify ADLs within a semi free-living setting.
Methods:
A total of 99 participants performed a broad set of indoor and outdoor activities. Frame-by-frame performance was compared to ground truth (direct observation) using overall accuracy, Cohen's Kappa, precision, recall, F1-scores and a normalised confusion matrix. Agreement on total activity duration was assessed using mean absolute percentage error (MAPE) scores and Bland-Altman plots.
Results:
Activity classification showed almost perfect agreement (mean overall accuracy 0.87, mean Cohen's Kappa 0.84). Excellent performance (F1 > 0.90) was achieved for sitting, walking with crutches and cycling, while standing, level and non-level walking showed good performance (F1 > 0.80). Most misclassifications occurred between level walking, hill walking and stairs. Duration-based analysis confirmed high accuracy for sitting, walking with crutches and cycling (MAPE ≤ 10%). Bland-Altman plots indicated overestimation of level walking and underestimation of hill and stair walking. Step count was highly accurate (MAPE < 5%), whereas stair count showed only reasonable accuracy (MAPE ≈ 26%).
Conclusion:
These findings demonstrate the system's strong potential for real-world monitoring and classification of ADLs while also highlighting the need for improved detection of non-level walking activities.
