Background: Activity monitoring with consumer- or research-grade devices has been used to evaluate gait and physical activity in neuromuscular disease. Wrist-worn (W) and smartphone-embedded (S) sensors are the most widely used sensing modalities. Instrumented insoles (I) have also become available to researchers in recent years. However, little is known about how classification performance is affected by the combination of these sensing modalities.
Objective: We compared the accuracy of 7 activity classification models, each relying on a combination of three, two, or a single sensing modality (W, S, I), under unstructured conditions (UC) and free-living conditions (FC).
Methods: After donning the sensors, 11 healthy adults performed 6 activities (sitting, standing, walking, stair descending, stair ascending, and jogging) first following a prescribed sequence (structured condition, SC), then in a self-selected order (UC). In a separate day, they wore the sensors while carrying out their usual daily activities (FC). Direct observation and a validated activity tracker were used for data labelling in SC/UC and FC, respectively. For each combination of sensing modalities, a Support Vector Machine classifier was trained with an optimized subset of input features extracted from the sensors. The target classes were 6 (i.e., the activities outlined above) and 3 (sedentary, standing, stepping) for UC and FC, respectively. Multi-session leave-one-out cross-validation was implemented to evaluate classification accuracy. Paired t-tests were carried out to check for significant (??=0.05) differences in accuracy across the sensing modalities.
Results: Among the single sensing modalities, the wrist (W) yielded the worst accuracy (UC: 70.9%, FC: 72.3%). For UC, the combination I-W-S demonstrated the highest accuracy (97.6%), however any other modality featuring the insoles (I-S, I-W, I) yielded comparable accuracy (? 95.2%). For FC, the combination W-S produced the best accuracy (96.4%), but any other combination featuring I or S resulted in comparable accuracy (? 93.9%).
Conclusions: Sensors attached to the lower extremities may lead to better accuracy in evaluating gait and physical activity and hence should be preferred to wristband-based sensors. Future work will include integrating the activity classification models with learning-based gait analysis methods previously developed by our group, to enable real-life gait analysis in persons with neuromuscular disease.