Development and validation of accelerometer-based activity classification algorithms for older adults: A machine learning approach

• Main Author: Sasaki, Jeffer Eidi
• Language: ENG
• Published: ScholarWorks@UMass Amherst 2014
• Subjects: Gerontology|Kinesiology|Bioinformatics
• Online Access: https://scholarworks.umass.edu/dissertations/AAI3615445

Machine learning algorithms to classify activity type from wearable accelerometers are important to improve our understanding of the relationship between physical activity (PA) and risk for physical disability in older adults. Therefore, the main objective of this dissertation was to develop and evaluate machine learning algorithms to predict activity type and intensity in older adults from a commercially available accelerometer (ActiGraph GT3X+). In Study 1, we developed machine learning algorithms to classify activity type and intensity from raw accelerometer data in older adults. Thirty-five older adults performed an activity routine comprised of different activities (5 min/activity) while wearing three ActiGraph GT3X+ activity monitors (dominant hip, wrist, and ankle) and a portable metabolic system. Accelerometer and steady-state metabolic data were used to develop artificial neural network, random forest, and support vector machine algorithms (ANNLab, RF Lab, and SVMLab) to predict activity type and intensity in older adults using 20 s classification intervals. Classification accuracy of the models in detecting five activity categories ranged from 87% (ANN Lab hip, RFLab hip, and SVMLab hip) to 96% (SVM Lab wrist). The biases and root mean squared errors (RMSE) for predicted METs ranged from -0.01 MET (RMSE: 0.54 MET) for the RFLab wrist algorithm to 0.02 MET (RMSE: 0.67 MET) for the ANNLab hip algorithm. Study 2 evaluated the performance of the RFLab and SVM Lab algorithms for predicting activity type in free-living conditions. Fifteen participants from Study 1 were observed for 2-3 h in their free-living environment while wearing three ActiGraph GT3X+ activity monitors (dominant hip, wrist, and ankle). The RFLab and SVMLab algorithms were applied to hip, wrist, and ankle accelerometer data to classify five activity categories. Direct observation of activity type and duration served as criterion measures to evaluate percent correct classification rates of the algorithms. Correct classification rates ranged from 49% (SVMLab hip, SVMLab wrist, and RFLab wrist) to 55% (SVM Lab ankle). New RF and SVM algorithms were developed using free-living accelerometer data (RFFL and SVMFL) and different classification intervals were also applied. Correct classification of activity types for the RFFL and SVMFL ranged from 53% (SVMFL wrist, 5 s classification intervals) to 71% (SVMFL ankle, 30 s classification intervals). Overall correct classification rates of up to 76% (RFFL hip and RFFL ankle, 30 s classification intervals) were achieved when classifying only three activity categories. Our machine learning algorithms accurately predict activity type from accelerometer data in older adults under 'laboratory conditions' but not in free-living conditions. We were able to improve free-living classification accuracy using algorithms developed under free-living conditions. Further refinement of the algorithms is required for achieving sufficient accuracy in classifying activity type in free-living older adults.