Disorganization domain as a putative predictor of Treatment Resistant Schizophrenia (TRS) diagnosis: A machine learning approach

Abstract

Background: Treatment Resistant Schizophrenia (TRS) is the persistence of significant symptoms despite adequate antipsychotic treatment. Although consensus guidelines are available, this condition remains often unrecognized and an average delay of 4-9 years in the initiation of clozapine, the gold standard for the pharmacological treatment of TRS, has been reported. We aimed to determine through a machine learning approach which domain of the Positive and Negative Syndrome Scale (PANSS) 5-factor model was most associated with TRS.

Methods: In a cross-sectional design, 128 schizophrenia patients were classified as TRS (n = 58) or non-TRS (n = 60) after a structured retrospective-prospective analysis of treatment response. The random forest algorithm (RF) was trained to analyze the relationship between the presence/absence of TRS and PANSS-based psychopathological factor scores (positive, negative, disorganization, excitement, and emotional distress). As a complementary strategy to identify the variables most associated with the diagnosis of TRS, we included the variables selected by the RF algorithm in a multivariate logistic regression model.

Results: according to the RF model, patients with higher disorganization, positive, and excitement symptom scores were more likely to be classified as TRS. The model showed an accuracy of 67.19%, a sensitivity of 62.07%, and a specificity of 71.43%, with an area under the curve (AUC) of 76.56%. The multivariate model including disorganization, positive, and excitement factors showed that disorganization was the only factor significantly associated with TRS. Therefore, the disorganization factor was the variable most consistently associated with the diagnosis of TRS in our sample.

Keywords: Antipsychotics; Clozapine; Disorganization; Machine learning; Schizophrenia; Treatment resistant schizophrenia.