Institutes outside Greece
Information retrieval and machine learning for conducting systematic reviews
Abstract
dc:descriptionInnovative research studies provide alternative methods to minimise the time and cost associated with all SR (Systematic Review) stages which have critical importance on Evidence-Based Medicine. The time to complete a SR is typically between 2 and 6 months and the cost is estimated in tens of thousands of pounds. This thesis makes an effort to address a key challenge faced in SRs which eventually will lead to important reductions in the amount of process time and production cost. Specifically, this thesis supplements a tested proposal to enhance the SR initial retrieval stage of screening and ranking medical documents so that all the relevant abstracts of healthcare articles are found as soon as possible before the non-relevant with minimum computational effort. A new DL (Deep Learning) and IR (Information Retrieval) method is proposed and evaluated against a highly effective and custom IR method on CLEF systematic reviews. The proposed strategy promotes the usage of an optimised probabilistic retrieval model along with different DL models. My hypothesis is that using DL an evolutionary branch of Machine Learning will result in higher IR accuracy than depending solely on prediction scores from conventional IR models. If this objective is attained, it could lead conceivably to the reduction of human effort and time to assess public health SRs. For this purpose, my research starts with an empirical collection of research knowledge and data (CLEF collections) in the examined field. Furthermore, this thesis contains a description of my experimentation with different document indexing, ranking, query processing and development methods to analyse the proposed approach by modifying and extending existing open-source toolkits but also combining these to construct my proposed IR pipeline. The conclusions drawn from the results indicated that including a greater amount of fields to an indexer for medical articles is a decisive positive contributor for the retrieval effectiveness of an IR system. Also, incorporating post-processing query expansion techniques based on thesaurus query formulation enrichment boosts further the IR performance of the proposed IR system. However, the size of a processed query by itself is not solely sufficient to expect improved performance on CLEF topics. Finally, I find that the proposed DL IR strategy works. I compare it to the recently successful BM25+RM3 (IR) model and conclude that the suggested method accomplishes advanced retrieval performance in the initial ranking of the articles with CLEF datasets for the CLEF eHealth Technologically Assisted Reviews in Empirical Medicine Task.
Degree
thesis:*- Grantor dc:publisher
- Institutes outside Greece
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ioannidis, Alexandros
- Ιωαννίδης, Αλέξανδρος
Subjects
dc:subject × 12- Ανάκτηση πληροφοριών
- Μηχανική μάθηση
- Συστηματικές ανασκοπήσεις
- Information retrieval
- Machine learning
- Systematic reviews
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57277