A Comparison of Machine Learning Approaches to Activity Classification Using IMU Data Collected in a Community-Based Setting

Abstract

Highlights What are the main findings? Several classical machine learning models outperformed two deep learning models (as well as a chance-level “dummy” classifier) in classifying functional activities using both a small and an expanded set of features derived from an IMU dataset from four wearable sensors. What are the implications of the main findings? Functional activity classification using IMU data from bilateral wrist and ankle sensors is feasible in a less constrained, community-based setting. Multiple ML models performed with balanced accuracy >60% (far better than chance performance) despite a small, heterogeneous dataset and short (5 s) input windows, supporting their potential use in small-sample clinical populations. For small-sample human activity recognition (HAR), model choice, feature complexity, non-target (“null”) activity handling, and sensor layout should be evaluated comprehensively in heterogeneous datasets for real-world applications. Abstract Machine learning (ML) algorithms can be used to extract clinically meaningful information from movement data captured by inertial measurement units (IMUs), but many human activity recognition (HAR) pipelines are developed on large laboratory datasets that may not reflect small, heterogeneous, real-world samples. The purpose of this study is to systematically compare the accuracy of multiple ML models, feature sets (both simple and expanded), class balancing strategies, null and transition period handling techniques, and sensor configurations for recognizing a set of four everyday activities extracted from IMU time series data from an age-diverse population. Six ML classifiers were trained and tested: multilayer perceptron, random forest, k-nearest neighbors, logistic regressor, CatBoost, and gaussian naive bayes. These approaches were used in a pipeline with differing sampling techniques including the synthetic minority oversampling technique or random undersampling, and feature handling steps including principal component analysis or a Select-From-Model metatransformer. Additionally, two deep learning methods, DeepConvLSTM and ResGCNN, were trained and tested. Accuracy, precision, recall, and area under the receiver operating characteristic curve were compared to a dummy classifier as a benchmark approximation to chance performance. All pipelines performed better than the dummy classifier, with model accuracy ranging between 0.427 and 0.644. This study demonstrated the ability of several ML algorithms to properly recognize a set of functional activities using limited IMU data from both children and adults.