Your browser does not support JavaScript!

Home    Collections    Type of Work    Doctoral theses  

Doctoral theses

Current Record: 3 of 2119

Back to Results Previous page
Next page
Add to Basket
[Add to Basket]
Identifier 000436164
Title Artificial intelligence in medicine : A clinical decision -support framework based on machine learning , statistical mixed-effect modeling and defeasible reasoning for rheumatoid arthritis long-term prognosis under biologic therapy
Alternative Title Τεχνητή νοημοσύνη στην ιατρική
Author Γενιτσαρίδη, Ειρήνη
Thesis advisor Σιδηρόπουλος, Πρόδρομος
Reviewer Πλεξουσάκης, Δημήτριος
Μαριάς, Κωνσταντίνος
Abstract Artificial intelligence (AI), is state-of-art information technology that provides intelligent software frameworks, able to collect information, perform data analysis and implement appropriate actions, to meet the needs of various environments, mimicking human behavior in reasoning, learning and problem resolution. The application of AI in medicine enables the collection and analysis of medical information and the implementation of appropriate actions, to support disease prevention, diagnosis, therapy and prognosis. Clinical decision-support systems (CDS) are systems that support the clinicians’ decision-making process. CDS systems are able to perform complex medical data analyses, unravel medical data associations, simulate and enhance the medical reasoning process and support prognostic information which assists the identification of critical situations. The present thesis focuses on the development of a clinical decision-support system (CDS-RA) to support Rheumatoid Arthritis (RA) management and long-term prognosis under biologic therapy. The CDS-RA utilizes artificial intelligence methods to conduct advanced medical data analyses based on statistical mixed-effect models, machine learning and defeasible reasoning. An important objective of the CDS-RA is to provide prognostic functionality able to early predict and reason about the persistent disease level of a RA patient under biologic therapy. Persistent disease level (PDL) was defined for patients under biologic therapy as the same disease activity level (DAS28 within a specific range) for at least half of the 5-year clinical follow-up, cumulatively and irrespective of fluctuations. Three PDL patient groups were specified, the LDA (DAS28&le;3.2), MDA (3.2&<DAS28&le;5.1) and HDA (DAS28>5.1) groups, respectively. The thesis provides evidence on the clinical importance of early patient categorization into the PDL groups by analyzing their association with different long-term outcomes. Patients’ data required for the analyses were retrieved from the Greek nationwide multicenter registry HeRBT (Hellenic Registry of Biologic Therapies) of seven healthcare centers in Greece. Two patient outcomes were compared between the groups, (a) the 5-year functionality trajectories and (b) the serious adverse events (SAEs) at 5 years of biologic therapy. A multivariable mixed-effect model was developed based on patients’ 5-year functionality trajectories which showed that MDA was associated with worse 5-year functionality course than LDA group (+0.27 higher HAQ trajectory in MDA than LDA, p<0.0001) and also HDA was associated with even higher 5-year functionality limitation than the LDA group (+0.69 higher HAQ trajectory in HDA than LDA, p<0.0001). Similarly, SAEs were differentiated (0.2±0.48 in LDA, 0.5±0.96 in MDA and 0.89±1.7 in HDA; p<0.01). The CDS-RA system provides a functional service that depicts the differentiated 5-year group trajectories of functionality and serious adverse events. The MDA patient group that under biologic therapy neither improves nor deteriorates (outside moderate disease activity) for a significant amount of time, is an under-researched group in RA literature. Thus, the thesis also focused on the analysis of this group and in particular its internal heterogeneity. Specifically, MDA patients were sub-categorized into two subgroups of lower and higher MDA. A multivariable mixed-effect model was developed based on patients’ 5-year functionality trajectories which showed that the higher MDA subgroup was associated with worse 5-year functionality course than the lower MDA subgroup (+0.26 higher HAQ trajectory in higher-MDA, p<0.0001). Similarly, SAEs were differentiated (0.32 ±0.6 in lower MDA and 0.64 ±1.16 in higher MDA; p=0.038). The heterogeneity found between lower and higher MDA patients can assist future T2T strategies to tailor treatments for these subgroups in order to improve their outcomes. The CDS-RA system includes an AI Layer that supports a prognostic functional service of patient PDL group (LDA, MDA, and HDA) based on three policies that utilize different medical evidence sources in decreasing priority. The first policy of highest priority is based on long-term disease data when they exist for a specific patient. Specifically, the policy categorizes the patient into a PDL group when the patient’s long-term follow-up fulfills the criteria membership for a PDL group (LDA, MDA, and HDA) by definition. The second policy is based on clinician’s expert opinion for group membership when it is provided. The third policy is a predictive service developed in the CDS-RA system for early prediction of the patient PDL group when neither long-term patient data exists, nor can clinicians provide information on the long-term disease level course that a patient will develop. The prognostic policies of the AI Layer and their prioritization were expressed with defeasible logical rules and were loaded in a AI engine integrated in the CDS-RA system that supports rule-based reasoning in Defeasible Logic. The AI logical rule theory is accessible, reusable, configurable and extendable to support additional prioritized medical policies. The predictive service of the AI Layer utilizes early (first 6 to 9 months) patient data in order to provide a personalized prediction of the patient’s long-term persistent disease level (LDA, MDA, or HDA). Two multivariable logistic regression Machine Learning models were developed. The first model yielded, males (OR 0.38 for females, p=0.02), lower baseline disease activity (OR 0.42 for DAS28 per unit, p=0.001), lower baseline functionality (OR 0.3 for HAQ per unit, p=0.01) and lower first semester’s average disease activity (OR 0.2 for DAS28 per unit, p<0.001) as early predictors for LDA compared to other groups. The second model yielded, younger age (OR 1.04 per year, p=0.003), shorter disease duration (OR 2.65 for duration years<2, p=0.026), prednisolone initiation at baseline (OR 1.81, p=0.033), lower baseline disease activity (OR 0.56 for DAS28 per unit, p<0.001), lower baseline functionality (OR 0.21 for HAQ per unit, p&lllt;0.001) and lower first semester’s average disease activity (OR 0.42 for DAS28 per unit, p<0.001), first semester’s average functionality improvement compared to baseline (OR 2.89, p=0.002) and lower occurrence of first semester’s serious adverse events (OR 0.32 for SAEs count>0, p=0.047) as early predictors for MDA compared to HDA. Overall, CDS-RA provides a state-of-art AI technological environment with a wide range of functional services that is mobile compatible, supports RA patient data management over time, facilitates patient-clinician interaction and provides personalized prognostic information for the long-term outcome of RA patients under biologic therapy. The innovative functionality integrates seamlessly statistical multivariable mixed-effect modeling, machine learning predictive modeling and defeasible logical reasoning to provide valuable insights during the clinical-decision making process. CDS-RA is aimed to assist clinicians in the biologic treatment of RA patients in order to support improved patient outcomes.
Language English, Greek
Subject Ρευματοειδής αρθρίτιδα
Issue date 2020-12-17
Collection   School/Department--School of Medicine--Department of Medicine--Doctoral theses
  Type of Work--Doctoral theses
Permanent Link Bookmark and Share
Views 10

Digital Documents
No preview available

Download document
View document
Views : 2