Comparative validity of energy expenditure prediction algorithms using wearable devices for people with spinal cord injury

Abstract

Study Design Cross-sectional validation study. Objectives To conduct a literature search for existing energy expenditure (EE) predictive algorithms using ActiGraph activity monitors for manual wheelchairs users (MWUs) with spinal cord injury (SCI), and evaluate their validity using an out-of-sample dataset. Setting Research institution in Pittsburgh, USA. Methods A literature search resulted in five articles containing five sets of predictive equations using an ActiGraph activity monitor for MWUs with SCI. Out-of-sample data was collected from 29 MWUs with chronic SCI who were asked to follow an activity protocol while wearing an ActiGraph GT9X Link on the dominant wrist. They also wore a portable metabolic cart which provided the criterion measure for EE. The out-of-sample dataset was used to evaluate the validity of the five sets of EE predictive equations. Results None of the five sets of predictive equations demonstrated equivalence within 20% of the criterion measure based on an equivalence test. The mean absolute error for the five sets of predictive equations ranged from 0.87–6.41 kilocalories per minute (kcal·min −1 ) when compared with the criterion measure, and the intraclass correlation estimates ranged from 0.06–0.59. The range between the Bland-Altman upper and lower limits of agreement was from 4.70 kcal·min −1 -25.09 kcal·min −1 . Conclusions The existing EE predictive equations based on ActiGraph monitors for MWUs with SCI showed varied performance when compared with the criterion measure. Their accuracies may not be sufficient to support future clinical and research use. More work is needed to develop more accurate EE predictive equations for this population.

Preliminary Field Validity of Free-Living Energy Expenditure Estimation in Wheelchair Users with Spinal Cord Injury via Wearable Device-based Models

“…We developed a new RF model using a relatively large dataset (N = 78) with a signi cant amount of activity minutes (7,640 minutes) across 28 different types of structured and unstructured activities. Our RF model achieved an overall MAE of 0.59 kcal/min and a MAPE of 23.6% on the training dataset, while the existing models achieved an overall MAE from 0.87-6.41 kcal and MAPE from 31-37% on a subset of the training dataset [19].…”

“…The model yielded a wide LoA from − 450.5 to 805.7 kcal/day [18]. In a recent study by our group, we evaluated ve sets of custom prediction models for MWUs, which were based on the ActiGraph devices (ActiGraph, LLC., Pensacola, FL, USA), the commonly used tri-axial accelerometer-based researchgrade wearable devices [19]. The ve models were developed using PA data collected in lab settings with relatively small sample sizes ranging from 15 to 49 and were validated on the same dataset used to develop the model.…”

“…The ve models were developed using PA data collected in lab settings with relatively small sample sizes ranging from 15 to 49 and were validated on the same dataset used to develop the model. We evaluated their performance on an out-of-sample dataset collected in lab settings from 29 participants [19]. However, the eld validity of these models remained unknown due to the lack of free-living data.…”

Determining Physical Activity Characteristics From Wristband Data for Use in Automated Insulin Delivery Systems

“…Many consumer products provide good enough accuracy for tracking various daily activities and ascertaining summary health statistics [10]- [14], [25]- [30]. Several studies monitor PA using mobile phones and sensors embedded in clothing, and some studies compare the effects of the wearable sensor location on the accuracy of PA tracking [31]- [38].…”