A carregar...
Publicação
Combining Synthesis of Cardiorespiratory Signals and Artifacts with Deep Learning for Robust Vital Sign Estimation
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 10.85 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Healthcare has been remarkably morphing on the account of Big Data. As Machine Learning
(ML) consolidates its place in simpler clinical chores, more complex Deep Learning (DL) algorithms
have struggled to keep up, despite their superior capabilities. This is mainly attributed
to the need for large amounts of data for training, which the scientific community is unable to
satisfy.
The number of promising DL algorithms is considerable, although solutions directly targeting
the shortage of data lack. Currently, dynamical generative models are the best bet, but focus on
single, classical modalities and tend to complicate significantly with the amount of physiological
effects they can simulate.
This thesis aims at providing and validating a framework, specifically addressing the data
deficit in the scope of cardiorespiratory signals. Firstly, a multimodal statistical synthesizer was
designed to generate large, annotated artificial signals. By expressing data through coefficients of
pre-defined, fitted functions and describing their dependence with Gaussian copulas, inter- and
intra-modality associations were learned. Thereafter, new coefficients are sampled to generate
artificial, multimodal signals with the original physiological dynamics. Moreover, normal and
pathological beats along with artifacts were included by employing Markov models. Secondly,
a convolutional neural network (CNN) was conceived with a novel sensor-fusion architecture
and trained with synthesized data under real-world experimental conditions to evaluate how its
performance is affected.
Both the synthesizer and the CNN not only performed at state of the art level but also innovated
with multiple types of generated data and detection error improvements, respectively.
Cardiorespiratory data augmentation corrected performance drops when not enough data is available,
enhanced the CNN’s ability to perform on noisy signals and to carry out new tasks when
introduced to, otherwise unavailable, types of data. Ultimately, the framework was successfully
validated showing potential to leverage future DL research on Cardiology into clinical standards.
Descrição
Palavras-chave
Multimodal Cardiorespiratory Signals Sensor Fusion Gaussian Copula Deep Learning Convolutional Neural Network Data Augmentation