Microsystèmes de capteurs utilisant l’intelligence artificielle embarquée pour le domaine médical

Abstract

TinyML is an emerging field dedicated to the development of compact, computationally and energy-efficient machine learning models suitable for embedded devices. While large-scale models require extensive computational resources and incur a significant environmental footprint, TinyML offers local, fast, and secure solutions, paving the way for innovative applications in areas such as healthcare and sports monitoring. In this thesis, I present the development of an innovative system that integrates a printed sensor matrix directly onto a KN95 mask, designed to monitor in real time the evolution of humidity on its surface and, by extension, to analyze the breathing patterns of its wearers. A humidity-sensitive ink based on bismuth ferrite (BiFeO3) was screen-printed onto a silver circuit, which was then sewn onto the mask and connected to a data acquisition system (Keithley DAQ6510). A series of experiments, comprising several cycles of normal and deep breathing, allowed us to build a dataset. These data were preprocessed and classified using an unsupervised approach with the HDBSCAN clustering algorithm, which enabled the identification of three distinct classes : normal breathing, deep breathing, and irregular breathing. The detection of irregular breathing episodes in certain participants was subsequently validated using a neural network, achieving an accuracy of 99.4% on the test data. Finally, the model was converted from TensorFlow to TVM to optimize its performance and ensure compatibility with an Arduino Nano 33 BLE, resulting in an inference time of less than 1 ms. The obtained results demonstrate the feasibility of a reliable and autonomous TinyML solution for monitoring breathing patterns in an embedded environment.

Date	31 Mar 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sylvain G. Cloutier (Supervisor) & Fabrice Vaussenat (Co-supervisor)