Jury :

Summary:

ardiovascular disease is the leading cause of death in the world, accounting for 31.4% of deaths worldwide in 2022. With more than half of cardiovascular disease-related deaths occurring outside hospital, there is a need for a constant home monitoring system. These monitoring systems need to be autonomous, comfortable and consume little energy. Since electrocardigram signals are by nature sparse events over time, a non-uniform quantization scheme is suitable for encoding this signal. This allows data-dependent energy consumption to be concentrated during periods of activity. In this context, this thesis studies the features used to differentiate between arrhythmic and healthy heartbeats used in cardiology. Particular emphasis is placed on the implementation of the feature extraction process in a
event-driven system in order to take advantage of the sparse nature of the signal. In addition, a study of integrated and non-integrated classification methods is presented, focusing on the search for ways to implement artificial intelligence directly in the system.
Combining an event-based sampling scheme with on-chip artificial intelligence (AI) thus aims to achieve a system complexity several orders of magnitude lower than that of state-of-the-art systems. Most event-driven systems use time-related features to classify arrhythmias, but they often require a sampling clock, which runs counter to the advantages of event-driven systems. In this thesis, we investigate a system in which the temporal information contained in sampled input data is extracted without a regular clock, using successive delays that coarsely encode the time between event-driven samples.
Using the delay output and the output of a Level-Crossing Analog to Digital Converter, an event-driven token is generated for each sample, containing coarse information about the shape of the ECG signal. Using these tokens as input to a neural network, an average accuracy of 98% can be obtained using the MIT-BIH database. However, as training a neural network with unbalanced datasets introduces bias, a new classification system based on antidictionaries is introduced. It consists of models created from negative information on healthy heartbeats, enabling the detection of arrhythmic heartbeats by training systems based on the same information.
based on AI without introducing bias. Classification based on antidictionaries achieves a similar average accuracy of 98 %, demonstrating its ability to classify arrhythmias at a level equivalent to state-of-the-art neural networks. The FPGA-based system requires 99% less memory to store the classifier, 1,024 logical comparisons instead of 512 multiplication operations and 99% less accumulation compared with the state of the art. This translates into 25.1% less Look-Up Table and 34.5% less registers used on the FPGA compared to the state of the art and does not require the implementation of any RAM. As a result of these advantages, the overall classification accuracy has been reduced by 1.03 percentage points compared with the state-of-the-art, with a classification accuracy of 97.97%.

Abstract:

Cardiovascular diseases are the leading cause of death globally, representing 31.4% of all deaths in the world in 2022. With more than half of all cardiovascular disease-related deaths happening in settings outside the hospital, there is a need
for constant at-home monitoring to prevent them. Such long-term monitoring solutions necessitate autonomous, comfortable, and low-power consumption devices. Since electrocardiogram signals are constituted by nature by sparse events in time, a non-uniform
quantization scheme is well suited to code this signal. It will enable a data-dependent power consumption concentrated during the activity periods.
In this context, this thesis investigates features used to differentiate arrhythmic and healthy heartbeats used in cardiology. A particular focus is made on the implementation of the feature extraction process in an event-driven system to benefit from the sparse nature of the signal. In addition, a study of classification methods both integrated and non-integrated is presented focusing on finding ways to implement artificial intelligence directly in the system.
Combining an event-driven sampling scheme with on-chip artificial intelligence (AI) thus aims to achieve a system complexity that is lower by several orders of magnitude than state-of-the-art systems. Most event-driven systems utilize time-related features to clas-
sify arrhythmia, but they often require a sampling clock, which counteracts the benefits of event-driven systems.
In this thesis, we investigate a system where the time information contained in the sampled input data is extracted without a regular clock, using successive delays coarsely coding the time between event-driven samples.
Using the output of the delays and the output of a Level-Crossing Analog-to-Digital Converter, an event-driven token is generated for each sample, holding coarse information on the shape of the electrocardiogram signal.
Using those tokens as the input of a neural network, an average accuracy of 98% can be achieved using the MIT-BIH database. However, because training a neural network with unbalanced datasets introduces a bias, a novel classification scheme based on antidictionaries is introduced. It consists of patterns created from negative information about healthy heartbeats, allowing the detection of arrhythmic heartbeats by training AI-based systems without introducing a bias. Antidictionary-based classification achieves a similar average accuracy of 98%, showing its ability to classify arrhythmia on par with state-of-the-art neural networks.
The system implemented on FPGA necessitates 99% less memory to store the classifier, 1,024 logic comparisons instead of 512 multiplication operations, and 99% less accumulation com pared to the state of the art.
This results in 25.1% less LUT and 34.5% less register used on the FPGA compared to the state of the art and does not require the implementation of any RAM. For those benefits, the overall classification accuracy was reduced by 1.03 percentage points compared to the state of the art with a classification accuracy of 97.97%.