The Ghost in the Machine : Machine learning models of the brain and genome in patients with schizophrenia and bipolar disorder

Abstract

The studies in this thesis used machine learning to explore brain abnormalities and genetic variation in patients with schizophrenia, bipolar patients and healthy controls. In Chapter 2, we studied the generalizability and possible clinical application to predict schizophrenia by structural magnetic resonance images of the brain. We included two large independent samples, one sample included 128 schizophrenia patients and 111 controls and was utilized to build a support-vector-machine model. The validation sample included 155 schizophrenia patients and 122 controls. We demonstrated that it is possible to achieve similar classification accuracy 71.4% and 70.4% respectively. Patients had relatively larger gray matter densities in the basal ganglia and left occipital lobe; and relatively small densities in frontal and superior temporal lobes and hippocampus. In Chapter 3, we created three models, each one separating two groups. The discovery sample included N=66 per group and the validation sample included N=46, 47, 43 schizophrenia patients, bipolar patients, and healthy controls, respectively. We confirmed the possibility to separate schizophrenia patients and controls with an accuracy of 75,5%. Moreover, patients with bipolar disorder and schizophrenia could be separated with an accuracy of 65,5%. The difference between two groups was apparent even though the samples were acquired on different scanners with different field strengths. Bipolar patients and healthy subjects could not be classified significantly above chance. In Chapter 4, we acquired data from recent onset psychotic patients. All patients underwent a baseline MRI scan and were assessed clinically 3 to 7 years thereafter. We included five samples from the University Medical Center Utrecht (n=67), the Institute of Psychiatry, Psychology and Neuroscience, London (n=97), the University of São Paulo (n=64), the University of Cantabria (n=107), and the University of Melbourne (n=54). Combining multi-center neuroimaging data led to a single classifying model that performs 90% accurate when classifying gender. Moreover, the results showed that a multicenter model improved the performance in individual smaller and possibly more heterogeneous single-center samples. When classifying illness course, defined as “continuous” illness course (no remission of symptoms of greater than 6 months); or “remitting” illness course (one or more periods of remission of at least 6 months, and no episode lasting longer than 6 months), classification accuracies ranged from below chance to 70%.

In chapter 5 we explored common genetic variations, i.e., single nucleotide polymorphisms (SNPs). We created three classification models based on a sample of schizophrenia patients and control subjects (N=705 cases and N=637 controls). The first model included all available SNPs (> 7 million SNPs). The second model included SNPs that had the best predictive value in the polygenic risk score (>700.000 SNPs) (Ripke et al., 2014). The final model included 74 independent GWAS-significant SNPs (Schizophrenia Working Group of the Psychiatric Genomics Consortium, 2014). The accuracies of these models were 54%, 60% and 56% respectively. Sensitivity and specificity were not adequate to be of diagnostic value. However, with much larger samples including all SNPs, potentially interesting differentiating SNP-patterns could be revealed.