Isabel Valera Martínez, a PhD student in the Signal Processing Group of the University Carlos III de Madrid will defend her doctoral thesis titled “Bayesian Nonparametric Modeling of Psychiatric Disorders” on November 21st
- Title: Bayesian Nonparametric Modeling of Psychiatric Disorders.
- Advisor: Fernando Pérez Cruz.
- Event Date: Friday, November 21, 2014, 11:00 am.
- Location: Adoración de Miguel (1.2.C16); Agustín de Betancourt Building; Leganés Campus; Universidad Carlos III de Madrid.
Mental health care has become one of the major priorities in developed countries, where the annual budgets assigned to mental health care reach hundreds of billion of dollars. Due to lack of laboratory tests as objective diagnostic criteria, there is not consensus among the psychiatrists either on the diagnostic criteria or the treatments. As a consequence, there exists an increasing interest in improving both the detection and treatment of mental disorders. This thesis is an interdisciplinary work, in which we study the causes behind suicide attempts and provide thorough analysis of pathological and comorbidity patterns of mental disorders. The final goal of this study is to help psychiatrists detect people with higher risk and guide them to improve treatments. To this end, we apply latent feature modeling to the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC), which collects information about the mental health of the U.S. population. In order to avoid the model selection step needed to infer the number of variables in the latent feature model, we make use of the Indian Buffet Process (IBP) . However, the discrete nature of the database does not allow us to use the standard Gaussian observation model, and therefore, we need to adapt the observation model to discrete random variables. In a first step, we propose an IBP model for categorical observations, which are the most common in the NESARC. We consider two likelihood observation models: a multinomial-logit and a multinomial-probit model. We derive efficient Monte-Carlo Markov chain (MCMC) inference algorithms that resort to either the Laplace approximation or the expectation propagation (EP) algorithm to compute the marginal likelihood. We also derive a variational inference algorithm that provides a less expensive, in terms of computational complexity, alternative to the samplers. Afterwards, to account for all the available information about the subjects (that includes also non categorical observations, such as age, incomes or education level), we extend the IBP observation model to handle mixed continuous (real-valued and positive real-valued) and discrete (categorical, ordinal and count) observations. This model keeps the properties of conjugate models and allows us to derive an inference algorithm that scales linearly with the number of observations. Finally, we present the experimental results obtained after applying the proposed models to the NESARC database, studying both the hidden causes behind suicide attempts and the pathological and comobidity patterns of mental disorders.