Learning to Recognise Exercises in the Self-Management of Low Back Pain

Globally, Low back pain (LBP) is one of the top three contributors to years lived with disability. Self-management with an active lifestyle and regular exercises is the cornerstone for preventing and managing LBP. Digital interventions are introduced in the recent past to reinforce self-management where they rely on self-reporting to keep track of the exercises performed. This data directly influence the recommendations made by the digital intervention thus accurate and reliable reporting is fundamental to the success of the intervention. In addition, performing exercises with precision is important where current systems are unable to provide the guidance required. The main challenge to implementing an end-to-end solution is the lack of public sensor-rich datasets to implement Machine Learning algorithms to perform Exercise Recognition (ExR) and qualitative analysis. Accordingly we introduce the ExR benchmark dataset “MEx”, which we share publicly to encourage future research. The dataset include 7 exercise classes, recorded with 30 users using 4 sensors. In this paper we benchmark state-of-the-art classification algorithms with deep and shallow architectures on each sensor and achieve performances 90.2%, 63.4%, 87.2% and 74.1% respectively for the pressure mat, the depth camera, the thigh accelerometer and the wrist accelerometer. We recognise the scope of each sensor in capturing exercise movements with confusion matrices and highlight the most suitable sensors for deployment considering performance vs. obtrusiveness.