Activity contextualisation and terrain classification via wearables in Parkinson’s

Engin F. V., Kuduz H., Kaçar F., Pearson-Noseworthy L. T., Das J., Stuart S., ...Daha Fazla

COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE, sa.26, ss.371-378, 2026 (SCI-Expanded, Scopus)

Özet

Background: Mobility impairment is influenced by intrinsic Parkinson’s disease (PD) factors but also by extrinsic/environmental factors, such as indoor vs. outdoor locations as well as terrain type. Most existing studies address human activity recognition (HAR) and terrain classification separately and often focus on healthy cohorts. There remains a need for an integrated framework that enables contextual mobility assessment in people with PD (PwPD) using wearable sensors.

Methods: In this exploratory pilot study, we developed a unified multimodal wearable framework based on a one-dimensional convolutional neural network (1D-CNN) to perform HAR and terrain recognition in PwPD. A local dataset was collected from ten PwPD using synchronised inertial measurement units (IMUs) and surface electromyography (sEMG) sensors positioned on the lower back and lower limbs. The model was evaluated under three sensor configurations (M1, lower-back IMU; M2, individual IMU windows pooled across four lower-limb sensor locations; M3, corresponding IMU+sEMG windows pooled across the same locations). External benchmark evaluation was conducted using the UCI-HAR, WISDM and Uneven Walking Surface IMU datasets to assess the applicability of the same architecture across independent datasets.

Results: On the local dataset, HAR accuracy increased from 0.786 with M1 to 0.889 with M3, while terrain classification accuracy increased from 0.821 to 0.881. Participant-level analysis showed significant differences across configurations for activity accuracy (p = 0.032, W = 0.383), precision (p = 0.001, W = 0.753) and F1-score (p = 0.008, W = 0.531). For terrain classification, significant differences were observed for recall and F1-score (both p = 0.025, W = 0.370). When independently retrained on the external datasets, the same architectural design achieved accuracies of 0.976 on UCI-HAR, 0.981 on WISDM and 0.875 on the Uneven Walking Surface dataset, demonstrating the applicability of the same architecture across independent datasets.

Conclusion: The proposed framework supports both HAR and terrain classification in PwPD using a common architecture and provides a basis for combining activity and environmental context in wearable mobility assessment. The findings provide preliminary evidence of performance differences across sensor configurations.

Keywords: Parkinson’s disease, human activity recognition, terrain classification, wearable sensors, one- dimensional convolutional neural network, CNN.