On the application of Machine Learning and Complex Networks to Neuroscience

Predictive models using complex network analysis of EEG and fMRI data can automatically detect brain changes associated with schizophrenia, Alzheimer's disease, and autism, and also identify functional changes induced by psychedelics such as ayahuasca and N,N-dimethyltryptamine. The models outperformed those previously reported in the literature. Interpreting the models and network metrics provided new insights into how these substances work. New metrics developed by the author were especially relevant for autism and psychedelic studies. The approach is flexible and likely applicable to other conditions and substances, like antidepressants.

Study at a glance

Abstract

Data mining and knowledge discovery is a research area with applications in various fields, such as medicine. Data mining methods have proven to be very effective in making automated diagnoses and help medical teams in decision making. In addition to using data mining, medical data can be represented by complex networks. In the case of the brain, for example, brain regions can be represented as vertices of a graph and the neural activity between the regions define the connection. In this way, we can compare the brain structure of healthy patients with that of patients with mental disorders to define diagnostic methods and gain insights into how brain structure is related to behavioral and neurological changes. The aim of the present work is to develop a predictive model that can improve the diagnosis of mental disorders such as schizophrenia, Alzheimers disease, and autism using electroencephalogram and functional magnetic resonance imaging data. In addition, it is to be tested whether the same workflow is capable of automatically detecting the influence of neurally active substances on functional changes in network structure. Because psychedelics are thought to have therapeutic potential for some mental disorders, data from experiments with ayahuasca and N,N-dimethyltryptamine were considered as examples. In general, the predictive models developed for the diseases were not only able to automatically detect the functional changes, but were also superior to the models presented in the literature. Regarding the investigation of psychedelics, it could be shown that the same workflow is equally suitable to automatically detect functional changes. Furthermore, by interpreting the models and metrics, new insights into the mechanisms of action of the substances could be gained. In addition, the present work determined which complex network measures are most effective in detecting brain changes, including new metrics developed by the author. The new metrics proved to be relevant to the studies of autism and psychedelics. It is likely that the methodology used here can be applied to other diseases and substances (e.g., antidepressants) due to its flexibility and adaptability to EEG and fMRI time series data