Machine learning as a statistical tool in schizophrenia research
Appears in the following Collection
AbstractSchizophrenia is a heterogeneous and multi-factored disease. Investigation of the disorder could profit from statistical methods which can address multiple putative factors and large, complex datasets. Machine learning is a branch of statistical analysis which has specialized in developing such methods. This dissertation contains four investigations of schizophrenia, each highlighting a different aspect of how machine learning can address topical questions in schizophrenia research.
The first study, "Potential genetic variants in schizophrenia: A Bayesian analysis," tested 36 candidate genetic loci to identify those which associated with increased risk of schizophrenia. Genetic effect sizes are small, requiring large samples to detect. Yet certain potentially interesting genetic variants are rare, making collecting such samples difficult. Early selection of genes worth further pursuit can save much wasted time and effort. Six loci were indicated.
The second study, "Morphological correlates to cognitive dysfunction in schizophrenia as studied with Bayesian regression," compared a set of brain morphological measures to identify those which best explained cognitive skill scores. Measures included volumes of cortical, subcortical, and cerebellar structure selected to reflect conflicting models of the morphological substrates of cognition and cognitive deficit in schizophrenia. It found that subcortical and cerebellar structures better explained cognitive skill than cortical structures.
The third study, "Investigating possible subtypes of schizophrenia pa- tients and controls based on brain cortical thickness," searched for cortical regions which showed evidence of morphologically distinguishable subtypes. The clinical heterogeneity of schizophrenia suggests that many disease factors may lead to morphologically distinguishable subtypes in patients. The same method applied to a mixed sample of case and control subjects provided a non-parametric investigation of cortical thickness variation in the disease. Morphological subtypes were not found in the patients. One third of the cortex was found to have two distinguishable types when patients and healthy control subjects were examined together.
The fourth study, "Grey and white matter proportional relationships in the cerebellar vermis altered in schizophrenia," hypothesized that proportional relationships between grey and white matter tissue volumes in the vermis would be strong in healthy control subjects and weakened in patients, reflecting an optimum balance dictated by contrasting biological constraints and disturbed in the disease. This was found to be the case, suggesting an alternate model for vermis neuropathology in schizophrenia.
These studies show that machine learning can identify promising avenues for further exploration, discern among overlapping hypotheses, elucidate the structure of the data, and allow the formulation of novel hypotheses based on the structure of the data.
List of studies
I. H Hall, G Lawyer, A Sillén, EG Jönsson, I Agartz, L Terenius, S Arnborg. Potential genetic variants in schizophrenia: A Bayesian analysis. The World Journal of Biological Psychiatry 8(1):12-22, 2007.
II. G Lawyer, H Nyman, I Agartz, S Arnborg, EG Jönsson, G Sedvall, H Hall. Morphological correlates to cognitive dysfunction in schizophrenia as studied with Bayesian regression. BMC Psychiatry 6:31, 2006.
III. G Lawyer, R Nesvåg K Varnäs, A Frigessi, I Agartz. Investigating possible subtypes of schizophrenia patients and controls based on brain cortical thickness. Psychiatry Research: Neuroimaging [in press].
IV. G Lawyer, R Nesvåg K Varnäss, G Okugawa, I Agartz. Grey and white matter proportional relationships in the cerebellar vermis altered in schizophrenia. Under review.