Subgrouping and prediction in first-episode schizophrenia

​This study uses advanced machine learning methods to explore merged datasets from the CINS/CNSR cohorts. Our aims are to subgroup patients with schizophrenia and identify biomarkers that can predict outcome and medication effects on the individual patient level.

The study is performed in collaboration with the Technical University of Denmark. In this study we combine data from three cohorts of anti-psychotic naïve, first-episode schizophrenia patients all collected in CNSR/CINS.

All cohorts have been examined using cognitive tests, structural MRI, electrophysiology, as well as genetics. Unsupervised machine learning such as deep neural networks is used to process the data and find underlying hidden structures. These will then be used as input for the supervised machine learning where the outcome can be used as label. The same approach is used in each cohort. The models can then be compared to identify whether the models are similar and the compounds used in the different cohorts can be predicted by the same model.

Our aim is to identify objectively measureable, biologically valid, and clinically meaningful subgroups of patients with schizophrenia by the application of advanced unsupervised machine learning approaches on our comprehensive data set of baseline cognitive, psychophysiological, structural and functional MRI data, as well as genetic, and neurochemical brain imaging (MRS/PET/SPECT) data from the CINS cohorts of antipsychotic-naïve and UHR patients.

Further we aim to predict clinical, biological, functional, and societal long-term outcome for future individual first-episode patients and UHR individuals by the application of supervised machine learning approaches on the above mentioned multimodal (baseline) data sets combined with long-term outcome data from WP2. This is the first step in developing individualized treatment strategies instead of the 'trial-and-error' approach applied in current clinical practice.

Our vision is to use any identified biomarkers to predict outcome or medication effects on the individual level. This will provide the basis for our long term goal: To develop a clinical tool that can help the clinical psychiatrists to identify the optimal treatment for the individual patient.

PhD theses based on data from this project:

  • Martin Axelsen "A multi-modal modelling approach to schizophrenia. Supervised and Unsupervised machine learning methods for sub-typing and prediction" (expected 2018)


Selected articles based on data from this project:

  • Bak N, Hansen LK (2016) Data Driven Estimation of Imputation Error—A Strategy for Imputation with a Reject Option. PLoS ONE 11(10)