Latent Profile Analysis and Conversion to Psychosis: Characterizing Subgroups to Enhance Risk Prediction

Kristin M. Healey; David L. Penn; Diana Perkins; Scott W. Woods; Richard S. E. Keefe; Jean Addington


Schizophr Bull. 2018;44(2):286-296. 

In This Article

Abstract and Introduction


Background: Groups at clinical high risk (CHR) of developing psychosis are heterogeneous, composed of individuals with different clusters of symptoms. It is likely that there exist subgroups, each associated with different symptom constellations and probabilities of conversion.

Method: Present study used latent profile analysis (LPA) to ascertain subgroups in a combined sample of CHR (n = 171) and help-seeking controls (HSCs; n = 100; PREDICT study). Indicators in the LPA model included baseline Scale of Prodromal Symptoms (SOPS), Calgary Depression Scale for Schizophrenia (CDSS), and neurocognitive performance as measured by multiple instruments, including category instances (CAT). Subgroups were further characterized using covariates measuring demographic and clinical features.

Results: Three classes emerged: class 1 (mild, transition rate 5.6%), lowest SOPS and depression scores, intact neurocognitive performance; class 2 (paranoid-affective, transition rate 14.2%), highest suspiciousness, mild negative symptoms, moderate depression; and class 3 (negative-neurocognitive, transition rate 29.3%), highest negative symptoms, neurocognitive impairment, social cognitive impairment. Classes 2 and 3 evidenced poor social functioning.

Conclusions: Results support a subgroup approach to research, assessment, and treatment of help-seeking individuals. Class 3 may be an early risk stage of developing schizophrenia.


Individuals at clinical high risk (CHR) often present with a mixture of difficulties in addition to subthreshold psychotic symptoms, such as neurocognitive decline, premorbid dysfunction, and anxious/mood disorders.[1–4] Heterogeneity impedes research by obscuring potentially discrete subtypes, which hinders clinical research, evaluation, and treatment.

Latent subgroup models are a novel approach in explicating risk in CHR and are within a group of statistical methods known as latent variable mixture modeling (LVMM[5]). LVMM, such as latent profile analysis (LPA), aims to identify homogenous subgroups within heterogeneous cohorts, each with independent symptom constellations and differential associations with conversion and functional ability.[5,6] LVMM may improve accuracy in identifying who among the CHR group will subsequently convert to psychosis. Imaging studies have provided support for latent CHR subgroups, finding significant neurobiological heterogeneity in gray matter volume.[7]

LVMM has been applied in 2 CHR studies with mixed results.[8,9] Velthorst et al[8] used a modified latent class factor analysis to investigate symptom profiles of 288 CHR and unaffected control (UC) individuals. "At risk" and "healthy" classes emerged, but classification did not enhance prediction of conversion. Possible reasons for this include incorporation of UCs with limited variability and lack of diversity in predictive indices.

Valmaggia et al[9] applied a LVMM approach to a sample of 318 CHR individuals' ratings on the Comprehensive Assessment of the At-Risk Mental States.[10] A 4-class model emerged, each associated with different rates of transition to psychosis. The subgroup with the highest transition rate (class 4, 41.2%) was characterized by the highest symptom ratings, lowest overall functioning, and highest unemployment rate. Classes were best separated by differences in negative symptoms and social/role functioning, indicating that these variables are useful in determining risk. Thus, LVMM identified individuals with a specific constellation of negative symptoms and role impairments that were associated with a higher conversion rate.[9]

The present paper seeks to extend previous LVMM findings through application of LPA in a large group of prospectively identified CHR and help-seeking control (HSC) individuals and build upon Valmaggia et al's[9] model by incorporating measures of pre-morbid, social and role functioning, neurocognition, and social cognition. The aim is to enhance model validity by adding diagnostically relevant clinical and neurocognitive indicators and to further characterize latent subgroups with covariates. Supplementary Table 1 defines all acronyms included in the present article.