Automatic Feature Selection and Classification of Physical and Mental Load using Data from Wearable Sensors

Long-term monitoring of health is essential in many chronic conditions, but automatic monitoring is not yet utilized routinely with mental stress. Accelerometers, magnetometers, ECG, respiratory effort, skin temperature and pulse oximetry were used with 12 health volunteers in this study for monitoring 1) heavy mental load, 2) normal mental load, 3) walking, 4) running and 5) lying. Heavy mental load consisted of a 20-min IQ test and normal mental load was represented by reading a comic book. Automatic feature selection using sequential forward search was used to select the best features for classification of the five activities. Normalized heart rate, utilizing activity context information was found to be the most powerful feature for discriminating heavy mental load from normal. Classification accuracy for all 5 activities was 89% with a custom decision tree and with a k-nearest neighbor classifier and 85% with an artificial neural network.