Application of machine learning in neuroscience

点击翻译按钮获取中文摘要

Major depressive disorder (MDD) is a serious mental condition that often completely debilitates a client. MDD is typically treated with one of several currently available antidepressant medications. However, the response rate to any of these medications is only about 30%. Unfortunately, there are currently no means for a priori assessment of whether a specific person will respond to a particular medication. Thus, in prescribing a treatment for MDD, the psychiatrist must by necessity resort to a trial-and-error procedure. This can result in long delays before remission and significant stress on the health care system. In conjunction with collaborating psychiatrists, the applicant has developed a preliminary EEG-based machine learning (ML) methodology that can predict the response of a person to an SSRI medication (which is one of the classes of anti-depressant treatment) before the therapy begins. It is clear that such a capability, when fully developed, will vastly improve the treatment of MDD. However, before this system can be exploited in clinical applications, significant further development of the ML methodology is required. The objective of the proposed research is therefore to develop new ML methods that can reliably predict response, not just to the SSRI class as is currently the case, but to a wider variety of pharmacological therapies. This requires development of new high-performance ML methods that are specific to this application. To improve performance, better classifiers and better features are required. We propose using EEG-based brain-source localization methods in conjunction with information theoretic criteria to identify brain sources that are directly associated with MDD. Extracting features from these sources should result in features with increased salience. We further intend to investigate improved classifier structures which can optimally exploit the nonstationarity and nonlinearity inherent in the underlying prediction model.