A Small network of Consumer-Based wearable sensors Enables The identification of The neutral- and Tilted-Pelvis running techniques

The success of technique-based training in running can be compromised by inter-individual variability. In a recent study using lab-based motion capture, we identified two clusters of runners exhibiting distinct preferred techniques: the neutral- and tilted-pelvis clusters. Cluster-specific interventions may enhance training success but currently require laboratory equipment to allocate new runners to the clusters. The purpose of this study was to develop an algorithm to classify runners into the proposed groups using wearable sensor data. The 84 runners involved in the initial cluster determination completed 4 minutes of treadmill running at 12 km/h whilst wearing pressure insoles and 6 inertial measurement units placed on popular consumer-based sensor locations. For every possible sensor network, domain knowledge and comprehensive time series feature extraction were used to generate input features for classification. A logistic regression model with elastic net regularisation was then developed to classify runners into the two classes. The top performing models used 2–3 sensors and achieved testing accuracies and f1-scores ≥ 0.82, enabling confident allocation of new runners to the neutral- and tilted-pelvis techniques via widely accessible wearable sensors. The identified sensor networks were sparse and can be seamlessly integrated into a runner's standard kit, making it feasible for wearable sensor companies to embed these models into existing technologies. The presented methodology facilitates large scale cluster-specific training development and testing, and provides the means for these interventions to reach the wider running community via more affordable wearable sensors, bridging the gap between laboratory research and in-field application.