Development of a portable predictive model for the real-time estimation of Pulmonary Ventilation

Abstract Monitoring physical activities and training through an ensemble of vitals such as heart rate or energy expenditure is a fundamental activity that allows coaches to gain useful insights on the training sessions while reducing the risk of overtraining and injuries. Only recently the additional evaluation of respiratory vitals during workout sessions has been proven to contribute to an exhaustive assessment of the physical exercise. However, golden standards for measuring respiratory parameters, such as the gas transfer machine, suffer from some major drawbacks, being extremely invasive and impractical. Therefore in this work, developed in collaboration with the startup StepUp Solutions, active in the field of hardware development for sports and health monitoring, we want to highlight the importance of obtaining fairly similar information relying on a non-invasive approach based on the estimation of the Pulmonary Ventilation from chest expansion and contraction signals measured with the StepUp Air wearable device. The focus of this thesis is on building and deploying on the nRF52840 processor (32-bit ARM® Cortex®-M4FCPU running at 64 MHz) exploiting the Arduino Nano 33 BLE board, portable predictive models to provide a real-time estimate of the Pulmonary Ventilation based on chest expansion and contraction data acquired by the StepUp Solutions team. We propose and compare the feasibility of two main solutions to predict MV exploiting two different machine learning approaches: a simple and compact Fully Connected (FC) network operating on interpretable features, manually extracted from the raw data and physiological parameters, and a custom designed Residual Depthwise Separable Convolutional (RSDC) network based on learnable features directly over the raw sequence of chest expansion and contraction data. Furthermore, both models were designed keeping into account limitations and constraints strictly related to the deployment environment: memory footprint and inference time. Overall, the FC network outperforms the RDSC network both in terms of precision than in terms of performances, obtaining a Mean Absolute Percentage Error on the prediction around 10% with an extremely low storage requirement (5.75 Kb vs. 24 Kb) and nearly instantaneous predictions (2.8 ms vs. 400 ms). However, even if currently the most suitable model to be deployed on the StepUp Air device is the FC network, we obtained promising results also exploiting our RDSC network. In terms of future work, we believe that the RDSC network can also be a valid solution as we hypothesize that such model could benefit from a larger training dataset for better results. In addition, a model with such architecture would reduce significantly the amount of work spent in developing custom C code for manual feature extraction.