Metrological characterization of a wearable device and a machine learning algorithm for assessment of sport activities performances

This thesis focuses on the comprehensive metrological characterization of the K-AI wearable device developed by Ksport, concentrating on gait and cardiac activity parameters. By comparing the device's measurements with certified counterparts (Pressure insoles produced by Novel for the gait parameters and Zephyr BioHarness 3.0 for cardiac related parameters), the study aims to establish the device's reliability in biomechanical analysis. In gait analysis, the K-AI device demonstrates high accuracy and precision in measuring gait duration for both feet. The mean residuals for the left foot are 0 s with standard deviation of 0.07 s, and for the right foot, they are 0.1 s with standard deviation of 0.05 s. These results emphasize the device's potential significance in biomechanical analysis during walking and running activities. Concerning cardiac activity analysis, despite an ongoing investigation into a data transmission issue affecting detailed heartbeat-by-heartbeat analysis, the K-AI device shows high accuracy and precision in acquiring average heart rate (HR). The residuals for mean HR analysis are distributed narrowly, with an average value of 0.3 bpm and a standard deviation of almost 1 bpm. The study extends to machine learning algorithms for classifying athlete activity levels. The regression model predicts numerical values with an R² of 0.98, RMSE of 0.15, and MAE of 0.07. The classification model exhibits an overall accuracy of 78%, with precision, recall, and F1-score varying across activity levels. In conclusion, the findings position the K-AI wearable device as a reliable tool for biomechanical and health monitoring applications. The study underscores positive results in heart rate and activity analysis, opening new perspectives for the device's application in sports contexts. Future optimizations, including addressing the ECG signal acquisition issue and refining machine learning algorithms, are recommended.