Anna Thalea Hoogestraat1,2,#, Alexandra Illiger1,2,#, Jörg-Hendrik Bach2,4, Sarah Blum2,3,4
1 Medical Physics, Carl von Ossietzky University of Oldenburg, Oldenburg, Germany; 2 HörTech gGmbH, Marie-Curie-Str. 2, Oldenburg, Germany; 3 Neuropsychology Lab, Department of Psychology, European Medical School Carl von Ossietzky University of Oldenburg, Oldenburg, Germany; 4 Cluster of Excellence Hearing4all, Germany; # These authors contributed equally
Background: Human activity tracking has gained momentum in recent years driven by the development of powerful mobile devices with integrated movement sensors. Among these are ear-level devices that have proven to produce rich data regarding physical motion and vital parameters. It seems likely that functionalities of hearing devices and personal health systems will continue to merge in the near future. As of now, many research areas investigate the informational content of movement sensors in small, body-worn devices. With this project, we add to the growing fields of health monitoring and ambulatory assessment by exploring a method for activity detection using hearables manufactured by Dopple BV (NL).
Methods: Pilot data from two individuals were collected using a 3-axis accelerometer integrated in the hearables. We used research prototypes with direct access to raw sensor data. Data were streamed via Bluetooth to a PC for processing. Participants performed 12 activities for 5 minutes each. After outlier removal, balancing between classes and scaling, data were classified using a Naïve Bayes classifier implemented in sklearn using a 5-fold stratified cross-validation procedure.
Results: Using mean acceleration of all three channels computed on subsequent windows of 1 second, activities were detected with 72 % and 67 % accuracy on average for each participant. The averaged confusion matrix across folds shows that accuracy varies over activities, which might be explained by the different statistical properties of the activities. In addition, systematic confusion between similar activities may occur.
Conclusion: We found accelerometer data from sensors integrated in a hearable and worn in the ear canal to be suitable for the detection of human activities in our pilot sample. More work is needed to further investigate real-time capabilities of our approach, but also the generalisation and robustness of our machine learning approach.
Figure 1 A: Data streamed from the hearables to a PC running a recording script. B. Data from two different activities, recorded over 5 minutes each. C. Results from a classification using Naïve Bayes classification and a 5-fold cross-validation procedure.