Machine learning-based ability to classify psychosis and early stages of disease through parenting and attachment-related variables is associated with social cognition

Abstract

Background: Recent views posited that negative parenting and attachment insecurity can be considered as general environmental factors of vulnerability for psychosis, specifically for individuals diagnosed with psychosis (PSY). Furthermore, evidence highlighted a tight relationship between attachment style and social cognition abilities, a key PSY behavioral phenotype. The aim of this study is to generate a machine learning algorithm based on the perceived quality of parenting and attachment style-related features to discriminate between PSY and healthy controls (HC) and to investigate its ability to track PSY early stages and risk conditions, as well as its association with social cognition performance.

Methods: Perceived maternal and paternal parenting, as well as attachment anxiety and avoidance scores, were trained to separate 71 HC from 34 PSY (20 individuals diagnosed with schizophrenia + 14 diagnosed with bipolar disorder with psychotic manifestations) using support vector classification and repeated nested cross-validation. We then validated this model on independent datasets including individuals at the early stages of disease (ESD, i.e. first episode of psychosis or depression, or at-risk mental state for psychosis) and with familial high risk for PSY (FHR, i.e. having a first-degree relative suffering from psychosis). Then, we performed factorial analyses to test the group x classification rate interaction on emotion perception, social inference and managing of emotions abilities.

Results: The perceived parenting and attachment-based machine learning model discriminated PSY from HC with a Balanced Accuracy (BAC) of 72.2%. Slightly lower classification performance was measured in the ESD sample (HC-ESD BAC = 63.5%), while the model could not discriminate between FHR and HC (BAC = 44.2%). We observed a significant group x classification interaction in PSY and HC from the discovery sample on emotion perception and on the ability to manage emotions (both p = 0.02). The interaction on managing of emotion abilities was replicated in the ESD and HC validation sample (p = 0.03).

Conclusion: Our results suggest that parenting and attachment-related variables bear significant classification power when applied to both PSY and its early stages and are associated with variability in emotion processing. These variables could therefore be useful in psychosis early recognition programs aimed at softening the psychosis-associated disability.

Keywords: Adult attachment style; Bipolar disorder; Machine learning; Parental care; Parental overprotection; Risk for psychosis; Schizophrenia.

Fig. 1

Probability of each feature for being selected in the machine learning Cross-Validation framework. Score closer to 1 represent a higher probability of being selected for decision by the Support Vector Machine algorithm. Abbreviations: PBI = Parental Bonding Instrument; ECR = Experiences in Close Relationships

Fig. 2

Depicting of the group x classification rate interaction on FEIT percentage of total correct responses in the discovery sample. Abbreviations: FEIT = Facial Emotion Identification Test; PSY = Individuals diagnosed with Psychosis; HC = Healthy Controls

Fig. 3

Depicting of the group x classification rate interaction on MSCEIT total score in the discovery sample. Abbreviations: FEIT = Mayer Salovey Caruso Emotional Intelligence Test; PSY = Individuals diagnosed with Psychosis; HC = Healthy Controls

Fig. 4

Depicting of the group x classification rate interaction on MSCEIT total score in the validation clinical sample. Abbreviations: FEIT = Mayer Salovey Caruso Emotional Intelligence Test; ESD = Early Stages of Disease; HC = Healthy Controls