125. The use of combinatorial supervised machine learning and wearable sensors to objectively measure activities of daily living from the Oswestry Disability Index

Abstract

BACKGROUND CONTEXT Disease specific health status is measured using survey-based patient-reported outcomes (PROs) (eg, Oswestry Disability Index, ODI) including questions about walking, sitting, standing, traveling (driving) and sleeping (lying down). However, survey data are difficult to collect routinely and subject to multiple biases. Multiple sensor and rules-based algorithms have shown positive predictive values (PPV) greater than 85% for key functional metrics such as walking, jogging, sitting and lying down in a laboratory setting, suggesting the potential to assess measured function rather than reported function. The purpose of this study is to test the predictive capability of a single wearable sensor containing a high-end inertial measurement unit (IMU) for core elements of the ODI survey in a real-world setting. PURPOSE To determine the feasibility of using a single wearable sensor in a real-world environment to measure a subset of activities in the ODI. STUDY DESIGN/SETTING Technology feasibility. PATIENT SAMPLE Five healthy volunteers willing to wear a sensor and undergo a 60-minute activity protocol. OUTCOME MEASURES Sensor generated predictions of activity measured at the 5 ms boundary. METHODS Training data were generated from five healthy volunteers wearing a single navigation-grade IMU (SBG Systems Ellipse2-N) for 60 minutes in a real-world environment. Volunteers were asked to sit, stand, lie supine, drive and walk in an unscripted manner. Data were continuously recorded at 200Hz. A combinatorial learning approach utilized multinomial logistic regression (MLR) to classify posture and sleeping results before feeding into a layer of support-vector-machines (SVM). Twenty percent of the data were held out for testing. RESULTS A total of 1,989,878 data points were analyzed from five healthy volunteers. After the two layers (MLR followed by SVM), accuracy was found to be 98.6% across all activities. Positive predictive values were 98.4% for walking, 97.8% for sitting, 98.5% for driving, 100% for lying and 97.9% for standing. CONCLUSIONS The use of a high-end IMU combined with machine learning algorithms 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 for key patient-centric activities routinely considered reflective of health status. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.