Abstract:
Every year, about 4 million people die from respiratory diseases. While early prediction would
reduce this mortality rate, till now diagnosis is only done at hospitals involving costly diagnosis
resources and scarce healthcare professionals. Ideally, regular noninvasive breath analysis checkups
at home would allow us to anticipate medical consultation. Considering developing country’s
contexts, existing commercial portable diagnosis kits under proprietary licenses are expensive and
require internet connectivity to communicate with the remote server running their cloud prediction
analytics. Thanks to recent advances in open source edge AI frameworks, this study presents a
design of an offline portable kit locally embedding a tiny Machine Learning (TinyML) trained
model to predict respiratory diseases. Evaluated on an open dataset of Chronic Obstructive
Pulmonary Disease (COPD), the resulting real-time requirements of our edge AI model is 15.9 Kb
of ROM and 1.5 Kb RAM and performs the inference in 1 ms. In addition, the use of synthetic
exhaled breath data to train an Edge AI model in cases of low datasets was also evaluated with the
model performance giving accuracies similar to that based on actual datasets. Results also show
that the accuracy and peak memory for the model are affected by pre-processing, type of sensors,
and the number of sensors. In addition to early detection of respiratory diseases, the proposed
solution will be of great value in the process of mass collection of exhaled breath data to
complement synthetic data and the few breaths that are collected in healthcare facilities. This will
enable the training of efficient AI models for respiratory diseases. Last but not least, the research
results from this master thesis have been published in 3 IEEE scopus-indexed conferences.
