Objective: To recognize human activity from accelerometer signals using unsupervised learning

Background: Monitoring of patients using wearables during their activities of daily living (ADL) could aid the treatment of movement disorders. It could improve quality of care and thus eventually patient’s quality of life. Human activity recognition (HAR) is often necessary as a preliminary step to extract meaningful outcomes from the acquired signals. However, solutions targeting the general patient population may have reduced performance due to differences in disease characteristics, severity and patient characteristics such as age. Similarly, physical activity (PA) intensity assessment needs expensive calibration studies to set thresholds for each target group that help interpreting the actual PA intensity from raw data metrics. Van Kuppevelt et al.[1] recently explored the use of Hidden Semi-Markov Models (HSMM), a type of unsupervised machine learning, as an alternative to a priori set threshold-based methods currently used in PA intensity assessment. We evaluate the possibility of using a similar method for HAR and explore different locations and number of sensors to improve accuracy.

Method: Nine healthy participants (mean 22.2, SD 1.7 years old), were measured during simulated ADL using Inertial Measurement Units (IMU) attached to lower back, waist and sternum and both shanks, thighs, wrists and upper arms. Participants were videotaped and the video was manually labeled identifying four activities: sitting, lying, standing, walking. An HSMM was trained to identify participant activity for every five second window using four features: the magnitude and the 3D angles of the acceleration vector. To determine accuracy, we assigned the label overlapping most with the manual labels to each of the states returned from the HSMM.

Results: When using the data from three accelerometers – sternum, left and right thigh – the trained model had a total accuracy of 71%. The model could distinguish the basic postures and movements well, but performed worse on the recognition of ‘intermediate’ states, such as sitting comfortably on a couch (sitting-lying error) and slowly moving around the house (walking-standing error).

Conclusion: Unsupervised learning may offer a viable HAR solution, but further research is necessary with different patient groups to test the reliability and robustness of the solution.

References: [1] van Kuppevelt, D. et al. Segmenting accelerometer data from daily life with unsupervised machine learning. PLoS One 14, e0208692 (2019).

« Back to MDS Virtual Congress 2021

MDS Abstracts - https://www.mdsabstracts.org/abstract/unsupervised-human-activity-recognition-from-accelerometer-data/