View Item 
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        JavaScript is disabled for your browser. Some features of this site may not work without it.

        Browse

        All of UU Student Theses RepositoryBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

        Heterogeneity in psychiatric disorders: Using machine learning to predict development of mood disorders in bipolar offspring

        Thumbnail
        View/Open
        Bachelor thesis Suzan Stempher 3865010.pdf (2.341Mb)
        Publication date
        2017
        Author
        Stempher, S.D.
        Metadata
        Show full item record
        Summary
        Heterogeneity in psychiatric disorders complicates the application of machine learning techniques in psychiatry. Various causes in the brain can lead to the same mental disorder, partitioning the patient data into several groups. In a classification task, it may be impossible to separate the patients from the rest with a linear boundary. Several methods with potential to deal with heterogeneity in data were discussed and applied to simulated data with a heterogeneous patient group. Support vector machines (SVMs) with a radial-basis kernel and artificial neural networks (ANNs) were able to separate the groups well, using a non-linear separation boundary. As a proof of principle, these machine learning methods where then used in a regression task performed on real-world data, aiming to predict age based on brain volumes (n=501). Significant improvements compared to the linear method were found, with the best model (ANN) having a mean absolute error of 4.3 years. Finally, classification models were trained on a clinical data set of children of bipolar parents (n=140) to predict development of mood disorders and bipolar disorder in 12 years. The linear and non-linear machine learning approaches performed similarly; this may be caused by a lack of heterogeneity in this particular data set. Still, accuracies ranging from 71 to 77% in predicting mood disorders and 76 to 80% in predicting bipolar disorder were reached. With follow-up research on new data sets, these promising results may be improved in order to apply the prognosis models in practice.
        URI
        https://studenttheses.uu.nl/handle/20.500.12932/27476
        Collections
        • Theses
        Utrecht university logo