IMPROVING PREDICTIVE ANALYSIS WITH LONGITUDINAL DATA USING DEEP LEARNING MODELS.

BASIC INFORMATION

Ph.D. Student: Roula Kadhim
Advisor: Sebastián Ventura
Started on: February 2021
Keywords: Deep learning, Time series, Longitudinal, Survival Analysis

THESIS PROPOSAL

Longitudinal studies employ continuous or repeated measures to follow particular individuals over prolonged periods of time often years or decades. They are generally observational in nature, with quantitative and/or qualitative data being collected on any combination of exposures and outcomes, without any external influence being applied. This study type is particularly useful for evaluating the relationship between risk factors and the development of disease, and the outcomes of treatments over different lengths of time. Similarly, because data is collected for given individuals within a predefined group, appropriate statistical testing may be employed to analyse change over time for the group as a whole, or for particular individuals. In contrast, cross-sectional analysis is another study type that may analyse multiple variables at a given instance, but provides no information with regards to the influence of time on the variables measure being static by its very nature. It is thus generally less valid for examining cause-and-effect relationships. Nonetheless, cross-sectional studies require less time to be set up, and may be considered for preliminary evaluations of association prior to embarking on cumbersome longitudinal-type studies.

Deep neural networks are a family of computational models that have led to a dramatic improvement of the state of the art in several domains such as image, voice or text analysis. These methods provide a framework to model complex, non-linear interactions in large datasets, and are naturally suited to the analysis of hierarchical data such as, for instance, longitudinal data with the use of recurrent neural networks.

The main goal is to improve longitudinal data modelling by using deep learning methods and their applications to solve many different problems in real life. The working hypothesis is that deep learning is an excellent methodology to reach this objective, so our primary interest will be the improvement of deep learning methods for that purpose. The next goal is to test the validation of these methods in order to apply them in real-world applications in many different sensitive fields.