1250 - A comparison of machine learning approaches to activity classification using inertial measurement units in a community based setting

Objectives: We sought to compare the ability of several machine learning (ML) algorithms to classify everyday activities based on features extracted from inertial measurement units.

Design: Thirty nine community members (22 female; age range: 3-69 years) were recruited to complete a series of four functional activities (folding towels, stacking cups, standing and sitting from a chair, dancing) and a fifth "null activity" consisting of waiting for the next activity to start. Each activity was performed for approximately 1 minute, with short intervening transitions. Subjects were monitored by four MetaMotion R+ activity monitors, one placed on each wrist and ankle, which collected accelerometer and gyroscope data. Time domain features from 5 second windows were extracted, and the dataset was then split into train and test portions. Classifiers trained and tested included: multilayer perceptron, decision tree, bagged decision trees, k-nearest neighbors, logistic regressor, gaussian naive bayes, and a dummy classifier. Performance between classifiers was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and running time.

Results: All six ML classifiers outperformed the dummy classifier on all metrics. A multilayer perceptron classifier performed best, with accuracy 0.725, precision 0.722, recall 0.742, and AUC 0.907.  A gaussian naive bayes classifier offered the fastest training and classification time. Selectively removing sensors or limbs from the training dataset degraded performance of the multilayer perceptron, but not the logistic regression classifier. Confusion matrices showed misclassifications between the null activity and other activities were most common.

Conclusions: Machine learning methods represent an effective means of classifying functional activities. Using four activity monitor devices may provide improvements over fewer devices. Automated activity classification can help to better understand everyday motion patterns to direct at home therapies and to support lifestyle-based interventions.