Machine-learning models for activity class prediction : a comparative study of feature selection and classification algorithms
Chong, Joana; Tjurin, Petra; Niemelä, Maisa; Jämsä, Timo; Farrahi, Vahid (2021-06-24)
Joana Chong, Petra Tjurin, Maisa Niemelä, Timo Jämsä, Vahid Farrahi, Machine-learning models for activity class prediction: A comparative study of feature selection and classification algorithms, Gait & Posture, Volume 89, 2021, Pages 45-53, ISSN 0966-6362, https://doi.org/10.1016/j.gaitpost.2021.06.017
© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
https://creativecommons.org/licenses/by/4.0/
https://urn.fi/URN:NBN:fi-fe2021111154662
Tiivistelmä
Abstract
Purpose: Machine-learning (ML) approaches have been repeatedly coupled with raw accelerometry to classify physical activity classes, but the features required to optimize their predictive performance are still unknown. Our aim was to identify appropriate combination of feature subsets and prediction algorithms for activity class prediction from hip-based raw acceleration data.
Methods: The hip-based raw acceleration data collected from 27 participants was split into training (70 %) and validation (30 %) subsets. A total of 206 time- (TD) and frequencydomain (FD) features were extracted from 6-second non-overlapping windows of the signal. Feature selection was done using seven filter-based, two wrapper-based, and one embedded algorithm, and classification was performed with artificial neural network (ANN), support vector machine (SVM), and random forest (RF). For every combination between the feature selection method and the classifiers, the most appropriate feature subsets were found and used for model training within the training set. These models were then validated with the left-out validation set.
Results: The appropriate number of features for the ANN, SVM, and RF ranged from 20 to 45. Overall, the accuracy of all the three classifiers was higher when trained with feature subsets generated using filter-based methods compared with when they were trained with wrapper-based methods (range: 78.1 %–88 % vs. 66 %–83.5 %). TD features that reflect how signals vary around the mean, how they differ with one another, and how much and how often they change were more frequently selected via the feature selection methods.
Conclusions: A subset of TD features from raw accelerometry could be sufficient for ML-based activity classification if properly selected from different axes.
Kokoelmat
- Avoin saatavuus [32049]