Machine learning for predicting psychotic relapse at 2 years in schizophrenia in the national FACE-SZ cohort

Prog Neuropsychopharmacol Biol Psychiatry. 2019 Jun 8;92:8-18. doi: 10.1016/j.pnpbp.2018.12.005. Epub 2018 Dec 12.


Background: Predicting psychotic relapse is one of the major challenges in the daily care of schizophrenia.

Objectives: To determine the predictors of psychotic relapse and follow-up withdrawal in a non-selected national sample of stabilized community-dwelling SZ subjects with a machine learning approach.

Methods: Participants were consecutively included in the network of the FondaMental Expert Centers for Schizophrenia and received a thorough clinical and cognitive assessment, including recording of current treatment. Relapse was defined by at least one acute psychotic episode of at least 7 days, reported by the patient, her/his relatives or by the treating psychiatrist, within the 2-year follow-up. A classification and regression tree (CART) was used to construct a predictive decision tree of relapse and follow-up withdrawal.

Results: Overall, 549 patients were evaluated in the expert centers at baseline and 315 (57.4%) (mean age = 32.6 years, 24% female gender) were followed-up at 2 years. On the 315 patients who received a visit at 2 years, 125(39.7%) patients had experienced psychotic relapse at least once within the 2 years of follow-up. High anger (Buss&Perry subscore), high physical aggressiveness (Buss&Perry scale subscore), high lifetime number of hospitalization in psychiatry, low education level, and high positive symptomatology at baseline (PANSS positive subscore) were found to be the best predictors of relapse at 2 years, with a percentage of correct prediction of 63.8%, sensitivity 71.0% and specificity 44.8%. High PANSS excited score, illness duration <2 years, low Buss&Perry hostility score, high CTQ score, low premorbid IQ and low medication adherence (BARS) score were found to be the best predictors of follow-up withdrawal with a percentage of correct prediction of 52.4%, sensitivity 62%, specificity 38.7%.

Conclusion: Machine learning can help constructing predictive score. In the present sample, aggressiveness appears to be a good early warning sign of psychotic relapse and follow-up withdrawal and should be systematically assessed in SZ subjects. The other above-mentioned clinical variables may help clinicians to improve the prediction of psychotic relapse at 2 years.

Keywords: Aggressiveness; Machine learning; Prediction; Relapse; Schizophrenia.

Publication types

  • Multicenter Study
  • Research Support, Non-U.S. Gov't

MeSH terms

  • Adult
  • Aggression
  • Cohort Studies
  • Diagnosis, Computer-Assisted* / methods
  • Female
  • Follow-Up Studies
  • Humans
  • Machine Learning*
  • Male
  • Prognosis
  • Psychiatric Status Rating Scales
  • Psychotic Disorders / diagnosis*
  • Psychotic Disorders / psychology
  • Psychotic Disorders / therapy
  • Recurrence
  • Schizophrenia / diagnosis*
  • Schizophrenia / therapy
  • Schizophrenic Psychology
  • Sensitivity and Specificity