Prediction models of functional outcomes for individuals in the clinical high-risk state for psychosis or with recent-onset depression: a multimodal, multisite machine learning analysis

PRONIA Consortium

Research output: Contribution to journalArticlepeer-review

109 Citations (Scopus)


Importance: Social and occupational impairments contribute significantly to the burden of psychosis and depression. To date, no risk stratification tools exist to inform personalized preventive strategies targeting functional disability in the at-risk and early phases of these illnesses.

Objective: First, to determine whether social and role functioning can be predicted in patients with clinical high-risk states for psychosis (CHR) or recent-onset depression (ROD) using functional, imaging-based, and combined machine learning (ML) models. Second, to assess the models’ geographic, transdiagnostic and prognostic generalizability, and compare them against rater-based prognostication. Third, to explore sequential prognosis encompassing clinical and combined ML models.

Design: The multi-site PRONIA study which naturalistically follows CHR and ROD patients, patients with recent-onset psychosis, and healthy controls over 18 months.
Setting: Seven academic early recognition services located in 5 European countries.

Participants: 116 CHR and 120 ROD patients recruited between 02/2014 and 05/2016 and followed on average (SD) for 329 (142) days.

Main Outcomes and Measures: Performance and significance of prognostic models.

Results: ML correctly predicted the 1-year social functioning outcomes in 76.9%/66.2% of the CHR/ROD cases using functional baseline data, in 76.2%/65.0% using structural neuroimaging, and in 82.7%/70.3% by combining data domains. Occupational impairment was less predictable using functional and combined ML, and unpredictable using neuroimaging. While lower functioning before study inclusion was a transdiagnostic predictor of social impairment, the neuroanatomical models showed diagnostic specificity: medial prefrontal and temporo-parieto-occipital gray matter volume (GMV) reductions, as well as cerebellar and dorsolateral prefrontal GMV increments were predictive in the CHR group, while reduced medio-temporal and increased prefrontal-perisylvian GMV was predictive in ROD. Poor prognoses increased the risk for psychotic, depressive, and anxiety disorders at follow-up in the CHR but not in the ROD group. ML outperformed rater-based prognostication in both study groups. Adding neuroimaging ML to functional ML provided a 2 to 11-fold increase of prognostic certainty in uncertain cases of the clinical ML.

Conclusions and Relevance: Social disability may act as a proxy of various poor outcomes in the CHR state. Precision medicine tools sensitive for impaired recovery from this core dimension of mental illness could facilitate more effective therapeutic strategies.

Original languageEnglish
Pages (from-to)1156-1172
Number of pages17
JournalJAMA psychiatry
Issue number11
Early online date26 Sept 2018
Publication statusPublished - Nov 2018


  • ocial functioning
  • Outcome heterogeneity
  • Multimodal prediction
  • Prognostic models
  • Psychosis
  • Depression
  • Predictive analytics


Dive into the research topics of 'Prediction models of functional outcomes for individuals in the clinical high-risk state for psychosis or with recent-onset depression: a multimodal, multisite machine learning analysis'. Together they form a unique fingerprint.

Cite this