Abstract
Background: People with schizophrenia (SCZ) suffer from impaired cognitive abilities and these are associated with poor functional outcomes. Cognitive Remediation Therapy (CRT) has been shown effective in improving the cognitive deficits of SCZ. Because there is evidence for CRT treatment heterogeneity of outcomes, there is a need to identify CRT predictors of differential response using moderation analysis of high dimensional psychiatric data, which typically contain relatively large percentages of missingness. This will contribute to precision medicine treatment, understanding mechanism responsible of differential therapy responses, and better prognosis.Aims: The primary aim of this PhD consisted of developing a CRT precision medicine model, using computer intensive statistical learning methods able to deal with high dimensional psychiatric data containing large percentages of missingness in the predictors and smaller per-centages in the outcome. Secondary aims were overcoming the following problems: variable selection or measurement of variable importance in the model, multicollinearity and overfitting, and summarising commensurate outcomes in one latent outcome.
Methods: A simulation study comparing four statistical learning methods (Lasso, Elastic-net, Random Forests and Conditional Inference Random Forests) combined with two missing data imputation techniques (Multivariate Imputation using Chained Equations and MissForest) was run. The combined methods were assessed according to their optimism-corrected (via bootstrap internal validation) prediction accuracy and variable selection performance in differ-ent scenarios. The best method was chosen to develop a CRT precision medicine model using individual participant data from seven randomised controlled trials with approximately 400 pa-tients. Factor scores from a latent summary measure of cognitive commensurate outcomes, obtained via Factor Analysis, was used as the model dependent variable, to accommodate the above univariate statistical learning methods.
Results: In the simulations, the method combining MissForest imputation with Lasso was the best compromise between prediction accuracy and clinical interpretability. MissForest-Lasso was then used to develop an internally validated precision medicine model, which se-lected only a weak moderator of treatment response. The model was therefore mainly prog-nostic.
Conclusion: In future research, more modalities of data, such as genetics, OMICS and neuroimaging data, are recommended to successfully identify moderators of CRT success.
| Date of Award | 2018 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Daniel Stahl (Supervisor), Artemis Koukounari (Supervisor) & Matteo Cella (Supervisor) |