Minimal Sensor Configuration for Human Activity Recognition in Patients With Hip Osteoarthritis: Proof-of-Concept Study.

Background: Hip osteoarthritis (OA) impairs function during activities of daily living (ADL); however, objective functional assessment methods are lacking. Objectively measuring and characterizing functional deficits during ADL can help rehabilitation specialists target the activities in which patients experience the greatest difficulties. To characterize daily living function, ADL must first be accurately identified. Human activity recognition (HAR) models based on wearable sensor measurements, such as inertial measurement units (IMUs), have been implemented in individuals with various functional impairments but not yet in people with end-stage hip OA. In addition, the balance between HAR accuracy and wearability, which is crucial for patients' acceptance, remains unclear. Finally, to assess deficits of pathological populations, comparison with asymptomatic controls is necessary. Objective: This study aimed to evaluate the minimal accurate IMU configuration for HAR, from 8 to 1 IMU, in patients with hip OA and asymptomatic controls for mobility-related activities. Methods: Data from 20 patients and 9 controls were included. Participants completed a tour of ADL in the hospital vicinity while equipped with 8 IMUs. Activities included gait on flat ground, gait up and down a ramp, stair ascent/descent, turns, sitting up/down, and static sitting. Bidirectional gated recurrent unit models were trained to classify ADL based on 3D accelerations and angular velocities of multiple IMU configurations. Model accuracy was assessed on test sets of patient and control data using the Cohen Kappa( κ). Results: The resulting κ values were 0.95 for 8 IMUs, 0.93 for 4 IMUs, 0.90 for 2 IMUs, and 0.79 for 1 IMU. The accuracy was higher for controls than for patients. The best 2-IMU configurations were 2 shanks and 2 feet, and the best single-IMU configuration was the shank. Conclusions: Reducing the number of IMUs from 8 to 2 showed only a minimal decrease in κ, suggesting that minimal IMU setups could be as accurate as larger setups. Regarding gait detection, a single IMU was sufficient to reach very high accuracy (κ≥0.90). The single pelvic IMU showed the lowest accuracy for gait while still reaching high accuracy (κ>0.65). These results underline the potential of single-IMU or 2-IMU configurations to recognize mobility activities under semistandardized conditions. However, given the small sample size, these findings should be validated in larger and more heterogeneous hip OA cohorts before any clinical application can be considered.