Evaluation of neural networks to identify types of activity using accelerometers.

PURPOSE: To develop and evaluate two artificial neural network (ANN) models based on single-sensor accelerometer data and an ANN model based on the data of two accelerometers for the identification of types of physical activity in adults.

METHODS: Forty-nine subjects (21 men and 28 women; age range = 22-62 yr) performed a controlled sequence of activities: sitting, standing, using the stairs, and walking and cycling at two self-paced speeds. All subjects wore an ActiGraph accelerometer on the hip and the ankle. In the ANN models, the following accelerometer signal characteristics were used: 10th, 25th, 75th, and 90th percentiles, absolute deviation, coefficient of variability, and lag-one autocorrelation.

RESULTS: The model based on the hip accelerometer data and the model based on the ankle accelerometer data correctly classified the five activities 80.4% and 77.7% of the time, respectively, whereas the model based on the data from both sensors achieved a percentage of 83.0%. The hip model produced a better classification of the activities cycling, using the stairs, and sitting, whereas the ankle model was better able to correctly classify the activities walking and standing still. All three models often misclassified using the stairs and standing still. The accuracy of the models significantly decreased when a distinction was made between regular versus brisk walking or cycling and between going up and going down the stairs.

CONCLUSIONS: Relatively simple ANN models perform well in identifying the type but not the speed of the activity of adults from accelerometer data.