Identifying clinical clusters with distinct trajectories in first-episode psychosis through an unsupervised machine learning technique

Abstract

The extreme variability in symptom presentation reveals that individuals diagnosed with a first-episode psychosis (FEP) may encompass different sub-populations with potentially different illness courses and, hence, different treatment needs. Previous studies have shown that sociodemographic and family environment factors are associated with more unfavorable symptom trajectories. The aim of this study was to examine the dimensional structure of symptoms and to identify individuals’ trajectories at early stage of illness and potential risk factors associated with poor outcomes at follow-up in non-affective FEP. One hundred and forty-four non-affective FEP patients were assessed at baseline and at 2-year follow-up. A Principal component analysis has been conducted to identify dimensions, then an unsupervised machine learning technique (fuzzy clustering) was performed to identify clinical subgroups of patients. Six symptom factors were extracted (positive, negative, depressive, anxiety, disorganization and somatic/cognitive). Three distinct clinical clusters were determined at baseline: mild; negative and moderate; and positive and severe symptoms, and five at follow-up: minimal; mild; moderate; negative and depressive; and severe symptoms. Receiving a low-dose antipsychotic, having a more severe depressive symptomatology and a positive family history for psychiatric disorders were risk factors for poor recovery, whilst having a high cognitive reserve and better premorbid adjustment may confer a better prognosis. The current study provided a better understanding of the heterogeneous profile of FEP. Early identification of patients who could likely present poor outcomes may be an initial step for the development of targeted interventions to improve illness trajectories and preserve psychosocial functioning.