Development of deep learning based language models for generating medical predictions in the field of psychological diseases

The analysis of medical texts by artificial intelligence systems offers important contributions to the hospital management process, patient health monitoring, personalized medicine recommendations and recommendation systems that support physician decision-making. Detecting patients with psychological disorders at an early stage and ensuring that these patients are referred to the right unit is an issue that will provide significant gains both in the treatment process and in the hospital operation process. Although the awareness of psychological disorders is at a very low level today, it is known that these patients frequently apply to different medical units for treatment. In order to contribute to the solution of these problems, the primary goal of this thesis is to analyze medical texts consisting of patient-physician interviews with Deep Learning systems and to identify new patients with potential signs of psychological illness. In this thesis, language representation models have been developed to detect patterns of semantic similarity between the sentences of patients who have received psychological treatment in the past and the sentences of potential new patients. In addition, other medical departments that potential psychiatry patients frequently consult, and common complications of psychological disorders are explained with empirical findings. According to the literature review in the field, it is thought to be the first time that a single model can be used for the early detection of any psychological disorder independent of the disease. In the proposed method, an architecture that produces approximately 89% of the same prediction with physician judgements is presented. With the use of this method, apart from individual determinations, the determination of social concerns and the change in psychological disorders caused by these concerns according to countries were analyzed. Two different data augmentation methods are proposed to minimize the disadvantages of difficulties in accessing medical data and lack of labelled data in similar studies. Together with these methods, Deep Learning-based language models that provide over 90% success in the detection of psychological disorders have been developed.