IMU, sEMG, or their cross-correlation and temporal similarities: Which signal features detect lateral compensatory balance reactions more accurately?

BACKGROUND AND OBJECTIVE: Falls are the leading cause of fatal and non-fatal injuries among seniors worldwide. While laboratory evidence supports the view that impaired ability to execute compensatory balance responses (CBRs) is linked to an increased risk of falling, existing unsupervised fall risk assessment methods are mainly focused on detecting changes in spatio-temporal gait parameters over time rather than naturally-occurring CBR events. To address the gap in available methods, this paper compares the capability of machine learning-based models trained on the kinematic data from inertial measurement units (IMU) and surface electromyography (sEMG) features to detect lateral CBRs, to ultimately address detection of CBRs in free-living conditions. Moreover, we propose a novel "Hybrid" feature set, which considers cross-correlation and temporal similarities between the normalized kinematic and sEMG signals. METHODS: Focusing on frontal plane perturbations, a classifier to automatically: 1) detect lateral CBRs during normal gait, and 2) identify type (i.e., crossover, sidestep) using data from three wearable IMUs and 4 sEMG signals from the thigh (i.e., biceps femoris, rectus femoris) and lower leg muscles (i.e., gastrocnemious, tibialis anterior) was developed. In total, 600 trials (including 358 lateral CBRs) from 7 young, healthy adults were analyzed. The effects of feature type (IMU, sEMG, Hybrid) and sensor placement on the random forest-based classifier performance were investigated. RESULTS: CBR detection (i.e., CBR vs normal gait) accuracies (leave-one-subject-out cross validation) were 83.95% and 99.21% using sEMG-based and IMU-based features, respectively, which dropped to 72.17% and 84.83% for the multiclass identification (i.e., side-step vs cross-over vs normal gait) problem. Findings yielded shank as the best overall location for the multiclass problem, and chest as the most accurate for CBR detection. In general, adding sEMG and Hybrid features to IMUs yielded incremental improvements in CBR detection and type identification (87.03% leave-one-subject-out cross-validation for type identification). CONCLUSION: The findings of this study demonstrate that IMU-based features are favourable over sEMG and Hybrid features for the task of CBR detection, with incremental value for type identification. Evidence presented suggests that Hybrid features may increase performance for other wearable sensor applications (e.g. activity recognition systems).