Predicting symptoms in first-episode psychosis using individual specific connectivity networks: an analysis of the human connectome project for early psychosis (HCP-EP)

Alexander Moscicki, MD, MS

MGH/McLean Adult Psychiatry Residency Training Program – Resident
Moscicki_Alexander poster

Scientific Abstract

Background: Most functional parcellations of the human cortex have been derived from group level data. While these atlases illustrate general patterns of cortical organization at a population level, they do not describe the heterogeneous functional architecture of individuals. Wang et al 2020 used a technique that parcellates the cortex at the individual level to identify individual specific functional connectivity in participants with schizophrenia spectrum disorders. In a machine learning model, individual specific functional connections, stratified by domain and diagnosis, robustly predicted symptoms. Models using group level parcellations failed to predict symptoms, highlighting the increased sensitivity of individually derived functional connectivity.

Methods: This study aims to replicate the findings of Wang et al 2020 in an independent sample from the Human Connection Project for Early Psychosis, a cohort of 151 individuals with a psychotic disorder (onset within 3 years of enrollment), ages 16-35. We identified homologous cortical parcellations for each participant using the individualized parcellation algorithm in Wang et al 2020 applied to resting-state fMRI data, after which we constructed individualized functional connectomes for each participant. We estimated symptom scores using support vector machine for regression models trained on individual functional connectomes and clinical scores. To evaluate model performance, we examined correlations between estimated and observed symptom measurement scores.

Results: Machine learning models using liberal and stringent cross-validation procedures have so far largely failed to correlate with observed symptom scores. Correlations between estimated and observed positive and negative symptoms scores were not significant (p>0.05), except for positive symptoms scores in individuals with schizophrenia and schizoaffective disorder (r= 0.381, p=0.0002).

Conclusions: While machine learning models using individual specific functional connectivity predicted symptoms in a prior sample of participants with psychotic disorders, such models failed to do so in this smaller sample of individuals with early psychosis.

research Areas


Alex Moscicki MD MSc, Lauren Luther PhD, Daniel Reznik PhD, Yoonho Chung PhD, Crystal Blankenbaker MRes, Alan Brier MD, Martha E. Shenton PhD, Daphne J. Holt MD PhD, and Justin T. Baker MD PhD

Principal Investigator

Daphne J. Holt MD, PhD, and Justin T. Baker MD, PhD