Extending schizophrenia diagnostic model to predict schizotypy in first-degree relatives

Sunil Vasu Kalmady,Animesh Kumar Paul,Russell Greiner,Serdar M Dursun,Andrew J Greenshaw,Ganesan Venkatasubramanian,Venkataram Shivakumar,Rimjhim Agrawal,Janardhanan C Narayanaswamy,Anekal C Amaresha

doi:10.1038/s41537-020-00119-y

Abstract

Recently, we developed a machine-learning algorithm “EMPaSchiz” that learns, from a training set of schizophrenia patients and healthy individuals, a model that predicts if a novel individual has schizophrenia, based on features extracted from his/her resting-state functional magnetic resonance imaging. In this study, we apply this learned model to first-degree relatives of schizophrenia patients, who were found to not have active psychosis or schizophrenia. We observe that the participants that this model classified as schizophrenia patients had significantly higher “schizotypal personality scores” than those who were not. Further, the “EMPaSchiz probability score” for schizophrenia status was significantly correlated with schizotypal personality score. This demonstrates the potential of machine-learned diagnostic models to predict state-independent vulnerability, even when symptoms do not meet the full criteria for clinical diagnosis.

Highlights

Genetic inheritance plays a strong role in the etiology of schizophrenia, representing ~80% of the liability for the illness, based on numerous twin and adoption studies[1,2,3]
This study examines whether a schizophrenia diagnosis model, learned using schizophrenia and normal functional magnetic resonance imaging (MRI) data sets, can identify higher schizotypal scores in first-degree relatives without schizophrenia
The current study explores an alternative approach for predicting the degree of schizotypal expression in unaffected first-degree relatives of schizophrenia patients

Summary

INTRODUCTION

Genetic inheritance plays a strong role in the etiology of schizophrenia, representing ~80% of the liability for the illness, based on numerous twin and adoption studies[1,2,3]. Given the strong evidence for familial aggregation of higher schizotypy expression in SSD10, we hypothesize that the first-degree relatives who were predicted by the model to have “schizophrenia” status, i.e., false positives (FP) will have significantly higher schizotypal scores, versus those who are predicted as non-schizophrenia status, i.e., true negatives (TN) by machine learning. This model classified 14 out of 57 subjects as FP, whereas the remaining 43 were classified as TN, based on the default threshold level of schizophrenia prediction probability >0.5. Further application of this approach holds significant promise for exploring related and comorbid symptom clusters in psychiatry

METHODS

Findings

CODE AVAILABILITY