| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 609.76 KB | Adobe PDF |
Authors
Abstract(s)
Amyotrophic Lateral Sclerosis (ALS) is a neurodegenerative disease vastly known
for its rapid progression, usually leading to death within a few years, by respiratory failure. Since there is no cure currently known, the main objective is treatment in order to
improve symptoms and prolong survival. A treatment that is known to be effective, is
Non-invasive Ventilation (NIV), being capable of extending life expectancy and improving quality of life. Therefore it is imperative to administrate it preemptively. However
Amyotrophic Lateral Sclerosis (ALS) affects most muscles of the body, so there will exist
many clinical conditions that require some kind of treatment. In this work, we propose
two approaches that use deep learning to predict Amyotrophic Lateral Sclerosis (ALS)
disease progression, to know when a patient should be treated or not for a given clinical
endpoint. In the first approach, we use the various snapshots of the patients as input without taking into consideration the temporal dependence between the available features, so
instead we use each assessment of the patient as a single instance to feed the models. In
this approach, we propose the use of a MLP and a CNN, to predict the outcome for the
time windows available (90, 180 and 365 days), using the instances mentioned before and
perform class resampling on the training sets, using SMOTE on the minority class and
Random Undersampling on the majority class. In order to have more data to train the
models and have a balanced set, which helps achieving a better performing model on the
test set. In the second approach, we take the snapshots of the patients and group them
by the patient reference, and proceed to only used those that have length of 3 and 4. On
the sequences of length 3, we performed padding in order to have the same length as the
sequences of length 4, by simply taking the first instance of this sequence and use it as
the first and second assessment on the sequences. We proceed to use the sequences to
feed them into a LSTM model, and train for the several datasets and retrieve the scores
obtained. On this approach we do not perform any type of class resampling, as the first
approach does. The results obtained show some promising insights, on the first approach
in which we use the instances of the patients, the preprocessing performed was one of the
main factors to the great results obtained. On the second approach, which the temporal
sequences of length 3 and 4 are used, the results obtained are not as promissing as the first
approach, however there are still improvements that could be done.
Description
Tese de Mestrado, Engenharia Informática, 2022, Universidade de Lisboa, Faculdade de Ciências
Keywords
ALS Functional Rating Scale Data Imputation Deep Learning Esclerose Lateral Amiotrófica Ventilação não-invasiva Teses de mestrado - 2022