Enabling Deep Learning Models for Seizure Monitoring within Wearables Lifecycle
Continuous monitoring of various physiological conditions enables early detection of potential health issues. The monitoring offers real-time insights into patient well-being and rapid medical interventions. As a case study, in epilepsy, a prevalent neurological disorder, mobile health monitoring using wearable systems can significantly reduce the impact of this condition and its potential long-term effects.
The lifecycle of wearable systems can be split into three main phases: the design phase, where foundational designing constraints for the wearable hardware and software are laid; the AI-based training phase, during which the initial data collection of the wearable device takes place and the device transitions to smart functionality; and the large-scale deployment phase, where the wearable's user base is expanded and long-term monitoring of the patients is enabled. This thesis studies the entire lifecycle of wearable systems, identifying key challenges and concerns within this domain. Both software- and hardware-based solutions are proposed to address these issues, enhancing the functionality and efficacy of wearable technologies in healthcare settings.
In the design phase of wearable systems, a key concern arises as large-scale transformer models often fail to meet the constraints of wearable devices related to energy usage, runtime, and memory capacity. To tackle these issues, in this thesis, I introduce a tightly-coupled,
small-scale systolic array for accelerating transformers and employing software optimizations.
The next concern in the design phase of a wearable system is choosing the best combination of electrodes for signal acquisition. These systems must balance performance, energy consumption, and ergonomic constraints when selecting an optimal subset of electrodes from a larger set, a problem with exponential complexity. In contrast, this thesis proposes a computationally efficient approach to explore all combinations.
As the wearable lifecycle moves into the AI-based training phase, new challenges arise: the need for substantial labeled data for initial model training and further model updates. In this thesis, to address these issues, I introduce a few-shot learning method that reduces initial data collection and incorporates a prototypical updating mechanism to simplify the update process and reduce energy consumption.
With a trained model now operational in our wearable system, the lifecycle proceeds to the large-scale deployment phase. In this phase, the patient data transmitted to a central medical unit for further analysis often involves sensitive personal information, raising privacy concerns for wearable systems. In this thesis, I demonstrate the potential for re-identification of patients in a de-identified dataset. Then, synthetic seizure signals generated by generative adversarial networks are proposed as a solution instead of transferring the real signals.
Finally, in the large-scale deployment phase, the growing number of patients using wearable devices for continuous health monitoring introduces the challenge of manually labeling EEG signals since it is a time-consuming process requiring expert analysis. This thesis proposes a deep learning architecture called Maximum-Mean-Discrepancy Decoder for offline detection and automatic temporal localization of seizures in long EEG recordings to assist medical experts in signal labeling.
EPFL_TH10370.pdf
Main Document
openaccess
N/A
16.29 MB
Adobe PDF
2146bf9561679dd38407be036716d5cd