Open-Source Remote Gait Analysis: A Post-Surgery Patient Monitoring Application

Abstract

Critical to digital medicine is the promise of improved patient monitoring to allow assessment and personalized intervention to occur in real-time. Wearable sensor-enabled observation of physiological data in free-living conditions is integral to this vision. However, few open-source algorithms have been developed for analyzing and interpreting these data which slows development and the realization of digital medicine. There is clear need for open-source tools that analyze free-living wearable sensor data and particularly for gait analysis, which provides important biomarkers in multiple clinical populations. We present an open-source analytical platform for automated free-living gait analysis and use it to investigate a novel, multi-domain (accelerometer and electromyography) asymmetry measure for quantifying rehabilitation progress in patients recovering from surgical reconstruction of the anterior cruciate ligament (ACL). Asymmetry indices extracted from 41,893 strides were more strongly correlated (r = −0.87, p < 0.01) with recovery time than standard step counts (r = 0.25, p = 0.52) and significantly differed between patients 2- and 17-weeks post-op (p < 0.01, effect size: 2.20–2.96), and controls (p < 0.01, effect size: 1.74–4.20). Results point toward future use of this open-source platform for capturing rehabilitation progress and, more broadly, for free-living gait analysis.

Estimating Biomechanical Time-Series with Wearable Sensors: A Systematic Review of Machine Learning Techniques

“…The majority of studies (80%) developed subject-specific models and only 33% of studies explored task extrapolation. The latter may be less of a barrier to implementation since in practice task identification will likely be a part of the pipeline for automated analysis [91], in which case highly accurate activity classification models are required [92]. Thus, task specific models could be selected following task identification.…”

“…The use of sEMG for remote monitoring is less common than accelerometry and has been used primarily for quantifying indices of physical activity [94,95,96,97]. Recent efforts have estimated muscle activation time-series during walking using methods similar to those used to estimate muscle force using Hill-type muscle models [8,91,98]. This pre-processing step was used by several reviewed papers suggesting they may be practically deployed.…”

Wearables-Only Analysis of Muscle and Joint Mechanics: An EMG-Driven Approach

“…The foot-ground contact and GP muscle synergy function models assumed walking gait. This would not be problematic for remote monitoring as walking activity is identified in the processing pipeline [2]. However, analysis of other tasks would require development of other task-specific models.…”

“…REMOTE patient monitoring, enabled by advances in wearable technology and algorithms for human movement analysis, promises to improve the assessment and treatment of musculoskeletal disease [1]. Recent work quantifying stride-by-stride gait mechanics at segment-, joint-, and muscle-specific levels has shown that these variables may provide more sensitive measures of patient health than the more typical gross measures of physical activity [2], [3]. Despite these advances, many of the most clinically relevant variables have yet to be observed outside of controlled, laboratory environments.…”

Interpretable machine learning models for classifying low back pain status using functional physiological variables

“…Second, the variables used for model building were collected in a single session, which although reflected a typical clinical assessment scenario, may not reflect normal movement behaviour in daily living. With more advance wearable sensor technology emerging which allows remote biomechanics analysis [49], the methods employed presently can be exploited to yield statistical models which have greater ecological validity, and ultimately better personalized predictive accuracy. Third, we did not include age as a predictor during model building even though individuals with LBP and rmLBP were significantly older than controls.…”