Application of Neural Network algorithms in biomedical data

Postgraduate Thesis uoadl:3245315 28 Read counter

Unit:
Κατεύθυνση Βιοπληροφορική-Υπολογιστική Βιολογία
Library of the School of Science
Deposit date:
2022-11-21
Year:
2022
Author:
Kane Rediona
Supervisors info:
Ηρακλής Βαρλάμης, Αναπληρωτής Καθηγητής , Τμήμα Πληροφορικής και Τηλεματικής, Χαροκόπειο Πανεπιστήμιο
Original Title:
«Εφαρμογή αλγορίθμων Νευρωνικών Δικτύων σε βιοϊατρικά δεδομένα»
Languages:
Greek
Translated title:
Application of Neural Network algorithms in biomedical data
Summary:
The aim of this present thesis is to apply Machine Learning and Deep Learning techniques on the biomedical data derived from the HELIAD epidemiologic study, referring to neuropsychiatric disorders in the third age. To accomplish that, Neural Network algorithms were utilized, provided by the refined Keras and TensorFlow APIs, as well as the highly efficient scikit-learn library .

The application of Feedforward Neural Networks was proposed for the multiclass classification of two distinct variables; variable G1 referring to the varying severity of memory loss with 5 classes, and variable G21 for the classification of participants with a clinical diagnoses of dementia, comprising of 3 classes. Autoencoders which are traditionally considered as Neural Networks models were also implemented, as well as the SOM clustering algorithm, which falls into unsupervised training methods.

The HELIAD raw data had two main drawbacks, which were the ample amount of missing values, and the high dimensionality of features, especially considering the rather limited number of participants. In an effort to tackle these problems, the proposed workflow attempts to incorporate two distinct preprocessing strategies, consisting mainly of two imputation techniques, and hyper-parameter tuning via Grid Search implementation.

As far as dimensionality reduction is concerned, some features were manually removed following meticulous inspection, while Autoencoders were additionaly developed as a means to distinguish the most important and informative feaures.

Many models were exhaustively examined, investigating different hyperparameters and architectures, in order to achieve the most reasonable performance. The best fitted models were selected not only on the basis of their classification efficiency, but on the overall optimal methodology comprising of the necessary pre-processing steps.
Main subject category:
Science
Other subject categories:
Technology - Computer science
Keywords:
Bioinformatics, algorithms, Neural Networks, Machine Learning
Index:
Yes
Number of index pages:
3
Contains images:
Yes
Number of references:
71
Number of pages:
111
thesis_Rediona_Kane.pdf (3 MB) Open in new window