Prognostic research in psychiatry: towards a clinically-relevant prediction model for first episode psychosis

Leighton, Samuel P. (2024) Prognostic research in psychiatry: towards a clinically-relevant prediction model for first episode psychosis. PhD thesis, University of Glasgow.

Abstract

Background. Prognosis is the determination of risk of future health outcomes in people with a given health condition. The primary aim for my thesis was to conduct prognostic model research into first episode psychosis (FEP). The prognosis of people with FEP is poor in around half of those affected and difficult to predict in individuals. Prognostic prediction models to predict outcome in individuals could facilitate early intervention to change clinical trajectories and improve prognosis. As part of my primary aim, I sought to answer four research questions. 1) Is prediction of individual patient outcome possible in FEP using clinical variables? 2) Does prediction model performance remain robust at external validation? 3) Does prediction model performance improve with the addition of biologically relevant disease markers? 4) Does prediction model performance improve with the application of advanced machine learning classifiers compared to logistic regression? These questions are addressed in studies 1 to 3.

The secondary aim for my thesis was to test whether routinely collected electronic healthcare record data could be used for prognostic research in the National Health Service (NHS) in Greater Glasgow and Clyde (GG&C). The coronavirus pandemic delayed collection of routine data in FEP. I took the opportunity to examine this question in a more common area of psychiatric disease, delirium, in the hope that information from this would inform future prospective studies in FEP. Delirium is an important risk factor for subsequent dementia. However, the field lacks large studies with long-term follow-up of delirium in subjects initially free of dementia to clearly establish clinical trajectories. This formed study 4.

Study 1. This study aimed to conduct a systematic review of prognostic prediction models developed for predicting poor outcome in FEP. Thirteen studies reporting 31 prediction models across a range of clinical outcomes met criteria for inclusion. Eleven studies used logistic regression with clinical variables. External validation was carried out in four studies. Only one study assessed whether biologically relevant disease markers added value as predictors. Two studies used machine learning but did not provide enough information to allow comparison to logistic regression. Most studies had methodological flaws and the potential for prediction modelling in FEP is yet to be fully realised.

Study 3. This study assessed the potential for biologically relevant disease markers as predictor variables and compared advanced machine learning classifiers to logistic regression in 168 patients with FEP. The addition of a biological variable did not improve the performance of a logistic regression model built using clinical variables. It is possible that the usefulness of the biological variables for prediction was curtailed by the lack of a mechanistic link to the pathophysiology of psychosis thereby limiting their effect size. The naïve Bayes machine learning model was better than maximum likelihood estimation (MLE) but not elastic net logistic regression in terms of discrimination. However, for all models except MLE logistic regression there were problems with calibration.

Study 4. This study consisted of a retrospective cohort study of all patients over the age of 65 diagnosed with an episode of delirium who were initially dementia free at onset of delirium within NHS GG&C between 1996 and 2020 using routinely collected electronic healthcare record (EHR) data. 12949 patients with an incident episode of delirium were included and followed up for an average of 741 days. The estimated cumulative incidence of dementia was 31% by 5 years. The estimated cumulative incidence of the competing risk of death without dementia was 49.2% by 5 years. The cause-specific hazard of dementia was increased with higher levels of deprivation and also with advancing age from 65, plateauing and decreasing from age 90.

Conclusions. Systematic review of the literature showed that there is considerable potential for prognostic prediction modelling in FEP, but that most existing models have methodological flaws. Developing on this literature, my FEP prognostic prediction model can help to identify individual patients at increased risk of nonremission at initial clinical contact and showed robust external validation. However, this approach did not benefit from the addition of biologically relevant disease markers as predictor variables or the application of machine learning methods. Finally, I demonstrated the feasibility of using routinely collected EHR data from NHS GG&C for prognostic research into delirium and the risk of subsequent dementia. This will inform future prospective prognostic modelling studies of routinely collected data in FEP. Altogether, this thesis made several contributions to the growing body of clinical prognostic research in first episode psychosis and delirium. In particular, considerable progress has been made towards the deployment of a useable and informative clinical prediction model which will improve care for people with first episode psychosis.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: R Medicine > RC Internal medicine > RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry
• Colleges/Schools: College of Medical Veterinary and Life Sciences > School of Health & Wellbeing > Mental Health and Wellbeing
• Funder's Name: Chief Scientist Office (CSO)
• Supervisor's Name: Cavanagh, Professor Jonathan, Krishnadas, Dr. Rajeev, Deligianni, Dr. Fani and Rogers, Dr. Simon
• Date of Award: 2024
• Depositing User: Theses Team
• Unique ID: glathesis:2024-84360
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 18 Jun 2024 10:03
• Last Modified: 19 Jun 2024 14:11
• Thesis DOI: 10.5525/gla.thesis.84360
• URI: https://theses.gla.ac.uk/id/eprint/84360

Related URLs

• Article DOI
• Article DOI
• Article DOI