T66. PREDICTING TREATMENT RESISTANT SCHIZOPHRENIA AT FIRST-EPISODE OF PSYCHOSIS

• Published in: Schizophrenia bulletin 2019-04, Vol.45 (Supplement_2), p.S229-S230
• Main Authors: Smart, Sophie, Agbedjro, Deborah, Consortium, STRATA-G, Pardinas, Antonio, Walters, James, Stahl, Daniel, Murray, Robin, MacCabe, James
• Format: Article
• Language: English
• Subjects: Poster Session I
• ISSN: 0586-7614; 1745-1701
• DOI: 10.1093/schbul/sbz019.346

Abstract Background Within clinical services, there can be a delay of years before patients with psychosis, who do not respond to antipsychotic medication, receive a diagnosis of treatment resistant schizophrenia (TRS). There is a need to identify these patients earlier within the course of their illness and expedite their access to specialist treatment. The aim of this project was to use data, available when patients first present to clinical services, to predict TRS. Methods We used existing prospective first-episode psychosis studies from across Europe (STRATA-G Consortium; N=2274, TRS=370 (16.27%)) and tested two statistical models: the first to find associations with TRS and the second to identify predictors of TRS. After using existing literature to perform variable selection, we ran a logistic regression to identify variables associated with TRS within our sample. We included all variables within a ‘Least Absolute Shrinkage and regression Operator’ (LASSO) logistic regression, tuned by bootstrapping to identify predictors of TRS. We report the LASSO predictive model that corresponds to the penalty which returns the area under the curve (AUC) within one standard error (1SE) of the maximum AUC. Internal validation of the LASSO model performance was run via bootstrap optimism-correction. Results In our multivariable logistic regression model, younger age of onset and less years in education were significantly associated with the odds of being TRS. The predictive model selected younger age of onset, less years in education, basic education vs. higher education, presence of a lifetime relationship vs. absence of a lifetime relationship, smoking vs. not smoking, poorer functioning measured using the Global Assessment of Functioning (GAF), and mode of onset between one and six months vs. longer than six months, as predictors of TRS. For a risk threshold of 50%, the AUC was 0.776 and the model had an accuracy of 83.9% (95% confidence interval: 82.3%, 85.4%). The sensitivity (correct identification of TRS) was 2.70%, with a positive predictive value of 58.82%, and the specificity (correct identification of responders) was 99.63% with a negative predictive value 84.05%. Discussion Our logistic regression model replicated previously published work on factors associated with TRS. Our prediction model performed well, but, due to the small proportion of TRS cases in our sample, was more accurate at identifying responders than patients with TRS. The predictors sel