Multimodal Prediction of Psychosis in the Prospective MoBa Birth Cohort

Abstract

There is a need for improved early psychosis detection beyond the traditional clinical high-risk strategy. Using the Norwegian Mother, Father and Child cohort study, we examined the predictive ability of self-reported psychotic experiences (Community Assessment of Psychic Experiences; CAPE) at age 14, in addition to general mental health factors, parent and childhood psychiatric diagnoses, schizophrenia polygenic risk scores, and birth-related factors, to predict subsequent psychosis onset using three machine learning approaches for imbalanced data. We explored also a multimodal prediction framework. For unimodal classification, we observed best balanced accuracies with general mental health factors (67.27 ± 1.76%), and CAPE (65.95 ± 1.09%). Multimodal models improved classification accuracy (68.38 ± 2.16%). With validation and additional model refinement, these features may be useful for initial screening within clinical stepped assessment frameworks.

Keywords: MoBa; machine learning; prediction; psychosis.

Figure 1

Balanced accuracies of the five unimodal prediction group models

Figure 2

Three individual predictor groups and a combined model

Figure 3

Steps for handling siblings within a stratified cross-validation framework Note. To make sure that siblings were always either in the training or test sets, we took a family-based sampling approach when dividing the data for cross-validation. 1) Individuals sharing the same parent IDs were defined as siblings. 2) We collapsed siblings into a single family unit with the case status being 1 if any of the siblings had a psychotic disorder. 3) We performed stratified cross-validation on the family unit, ensuring an equal number of cases and controls between training and test sets. 3) Once samples in each set were defined, we transformed the family units back into individual observations.