Mental disorders are complex phenomena that can be examined on multiple levels, including genetics, neural networks, cognitive processes, behavior, and phenomenology. However, contemporary psychiatric diagnostics primarily rely on descriptive criteria that focus on behavioral and phenomenological alterations, while disregarding neurobiological and cognitive deficits. This gap between neuroscience and clinical psychology has led to high comorbidity rates between disorders and limited success in treatment. To address this issue, novel research paradigms, such as computational psychiatry and machine learning, have emerged, offering potential solutions. In this study, we aimed to explore the potential of deep learning, a subtype of machine learning, in identifying and describing the neurobehavioral features of psychosis spectrum disorders. Specifically, we constructed deep learning models that utilized global brain connectivity measures to predict diagnosis and other psychopathological measures. We compared the performance of our models with those from previous studies, assessed the effectiveness of deep learning compared to canonical correlation analysis, and evaluated the success of deep learning models in predicting psychopathology based on both theoretical and data-driven indicators. Our sample consisted of 636 participants, including 150 diagnosed with bipolar disorder, 119 with schizoaffective disorder, 167 with schizophrenia, and 202 without diagnosis. We trained deep learning models for four tasks: binary and multiple classification of diagnosis, regression of data-driven indicators of psychopathology, and regression of theoretical measures of psychopathology based on global brain connectivity. The models were successful (98.34%) in differentiating between healthy and non-healthy participants, with important brain areas identified consistent with previous studies. However, accuracy in identifying specific psychotic disorders was lower (56.04%), and the internal processes of the models were less reliable. Comparing deep learning to canonical correlation analysis, deep learning models explained more variance in regression tasks. However, the models struggled to generalize to a test sample. Interestingly, deep learning models performed slightly better with theoretical measures of psychopathology compared to data-driven indicators. These findings highlight the potential of deep learning in neuroimaging data analysis, as it enables accurate differentiation between healthy and non-healthy individuals based on accessible data and captures complex neurobehavioral relationships that canonical correlation analysis may miss. Considering the convergence of the deep models' internal processes with previous findings on psychotic disorders, these models may hold practical value as screening tools. Nonetheless, the limitations of the deep learning models in more complex tasks underscore the need for further model development and larger datasets.
|
