Automatic diagnosis of mental disorders

A significant number of people worldwide suffer from mental disorders such as depression, bipolar and neurodegenerative disorders. These disorders adversely impact the life quality of people and have a significant economic impact on health-care providers. The World Health Organisation (WHO) considers depression as a common mental disorder, with more than 300 million people of all ages affected by depression. Similarly, Bipolar disorder affects more than 60 million individuals, which makes it among the most spread mental disorders worldwide. Even though the treatment of mental disorders has proved to be efficient in most situations, incorrect diagnosis is a common obstacle. This is because self-administered questionnaires and clinical interviews are the only available methods for diagnosis. These methods are influenced by subjective bias from clinicians or patients, time-consuming and hard to repeat. There is growing attention on early diagnosis of mental disorders as evolving treatments are expected to be more effective before irrevocable changes have occurred in the brain. The integration of novel methods based on the automatic analysis of visual signals may provide more information about a person's mental state, which could contribute to the clinical diagnostic process. This thesis demonstrates that eye features extracted from video recordings of patients' answers to a clinician's questions can help in the automated diagnosis of depression. Results show that there is eye blink abnormality among depressed patients due to psychomotor retardation. This manifests as longer eye blink duration. The efficacy of these features demonstrated in depression severity prediction where it achieved mean absolute error (MAE) of 8.30 and classification accuracy of 93% using the Audio/Visual Emotion Challenge 2014 (AVEC2014) dataset. Furthermore, visual features of eye movements combined with head pose features extracted from video recordings of patient's during a clinical interview in a specialist memory clinic for the development of an automated visual screening method to support the preliminary detection of patients with cognitive concerns related to progressive neurodegenerative disorders (ND), FunctionalMemory Disorder (FMD) mild Cognitive Impairment (MCI). Finally, a novel automatic diagnosis method for bipolar disorder developed based on deep learning. The proposed method investigated the application of several deep neural network architectures to extract the complex temporal trajectories of physiological behaviours that occur at multiple time scales. Results achieved shows that this model can be used as a universal automatic feature extractor for mental disorders. This is confirmed by testing the proposed model on AVEC2018 bipolar disorder dataset and AVEC2014 depression dataset. In the AVEC2018 dataset, individuals with bipolar disorder are classified into states of remission, hypo-mania and mania. The proposed model achieved Unweighted Average Recall (UAR) of 55.56%. Using the same model on AVEC2014 depression dataset, Mean Absolute Error (MAE) of 7.65 achieved which outperforms most of the previous studies on the same dataset. These results confirm the generalisability of the proposed deep learning model and that it can be used as a tool for multi-scale feature extraction.