Prediction of relapse in a French cohort of outpatients with schizophrenia (FACE-SZ): Prediction, not association.
Barbosa S., Tamouza R., Leboyer M., Aouizerate B., Andrieu C., Andre M., Boukouaci W., Capdevielle D., Chereau I., Kobayashi JC., Coulon N., Dorey J-M., Davidovic L., Dubertret C., Fakra E., Fond G., Goze T., Khalfallah O., Leignier S., Llorca PM., Mallet J., Martinuzzi E., Misdrahi D., Oriol N., Pignon B., Rey R., Roux P., Schürhoff F., Schorr B., Urbach M., Very E., Wu C-L., Benros M., Simon J., Hasan A., FondaMental Academic Centers of Expertise for Schizophrenia (FACE-SZ) collaborators None., Glaichenhaus N., Godin O.
BACKGROUND: Schizophrenia (SZ) commonly manifests through multiple relapses, each impeding the path to recovery and incurring personal and societal costs. Despite the identification of various risk factors associated to the risk of relapse, the development of accurate algorithms predictive of relapse has been limited, partly due to inadequate statistical methods. Additionally, despite the wealth of data showing strong associations between inflammation and schizophrenia, the two existing studies failed to demonstrate whether inflammatory parameters could predict relapse. Our goal is then to identify clinical and inflammatory parameters associated with relapse in schizophrenia and to develop model to predict relapse in each patient. METHODS: We have used classical Cox regression, survival penalized regression, as well as survival random forests to analyze clinical and inflammatory biological data collected in the network of the Schizophrenia Expert Centers in France in which individuals with SZ are clinically assessed and followed up annually for 3 years. RESULTS: Among 247 individuals with SZ, 71 (29 %) experienced a psychotic relapse during the 3-year follow-up period. The variables most consistently associated with relapses were smoking status, severity of positive symptoms and low global functioning. From a panel of inflammatory parameters, only IL-8 serum levels were associated with time to relapse. The predictive performance, assessed using C-index, was 0.54 using both penalized regression and random forests. CONCLUSIONS: We found several clinical and biological variables consistently associated with relapses across three distinct statistical methods. However, despite these associations, the predictive capacity of these models remained low, highlighting that association does not necessarily mean prediction.