Skip to main content
  • Machine Learning and Time Series Motion Classification: Is it possible to write real-world, single sensor algorithms to detect the activities of daily living used in patient reported outcome questionn

    Final Number:

    Nicholas M Benson PhD; Arjun Siby Kurian M. Eng.; Andrew Fields M. Eng.; Newton Metcalf; Wing Au MS; Seth Georgian BS; Domagoj Coric MD; Shane Burch; Steven D. Glassman MD

    Study Design:

    Subject Category:

    Meeting: Section on Disorders of the Spine and Peripheral Nerves Spine Summit 2019

    Introduction: Clinical improvements in function are often measured using survey-based PROs (e.g. ODI) with questions about walking, sitting, standing, traveling (driving) and sleeping (lying down). However, survey data is time consuming for both the patient and clinician. Multiple sensor and rules-based algorithms have shown positive predictive values greater than 85% for ability for walking, jogging, sitting, and lying down in a laboratory setting. The purpose of the current study is to determine positive predictive value of activity in a real-world setting of a single wearable sensor containing an extremely high-end inertial measurement unit (IMU).

    Methods: Three healthy volunteers wore a single navigation-grade IMU (SBG Systems Ellipse2-N) for 60 minutes in a real-world (outside) environment and asked to sit, stand, lie supine, drive and walk in an unscripted manner. Data were continuously recorded at 200Hz using a combinatorial supervised ML approach. A multinomial logistic regression (MLR) was first used to classify posture with results fed to a second layer of support-vector-machines (SVM). A classifier finalized the results in 1 second boundaries by rolling 5ms samples into a single vote. A 48-minute holdout data set was reserved for testing.

    Results: 573,600 5ms data points were analyzed. After the two layers (MLR followed by SVM), accuracy was found to be 99.65% across all activities. Positive predictive values were 99.24% for walking, 99.67% for sitting, 99.8% for driving and 100% for both lying and standing.

    Conclusions: The use of a high-end IMU and ML algorithms in this study to classify a subset of ODI activities in a real-world setting achieved greater accuracy than multiple sensors and rules-based algorithms reported in prior studies. This type of technology offers the potential for measured function to replace or supplement patient reported function focused on established patient-centric activities routinely considered reflective of health status.

    Patient Care: Paper-based patient reported outcomes are an imperfect way to measure health status. The use of noninvasive sensors has the potential to augment or replace the subjective outcome measures frequently used today.

    Learning Objectives: By the conclusion of this session, participants should be able to: 1) Consider the pros and cons of collecting objective activities of daily living using noninvasive sensors. 2) Discuss in small groups the strengths and weakness of this initial work and the potential for clinical studies.

    References: Lugade, Vipul, et al. "Validity of using tri-axial accelerometers to measure human movement—Part I: Posture and movement detection." Medical engineering & physics 36.2 (2014): 169-176.

We use cookies to improve the performance of our site, to analyze the traffic to our site, and to personalize your experience of the site. You can control cookies through your browser settings. Please find more information on the cookies used on our site. Privacy Policy