Mental and neuro-developmental disorders such as depression, bipolar disorder, and autism spectrum disorder (ASD) are critical healthcare issues which affect a large number of people. Depression, according to the World Health Organisation, is the largest cause of disability worldwide and affects more than 300 million people. Bipolar disorder affects more than 60 million individuals worldwide. ASD, meanwhile, affects more than 1 in 100 people in the UK. Not only do these disorders adversely affect the quality of life of affected individuals, they also have a significant economic impact. While brute-force approaches are potentially useful for learning new features which could be representative of these disorders, such approaches may not be best suited for developing robust screening methods. This is due to a myriad of confounding factors, such as the age, gender, cultural background, and socio-economic status, which can affect social signals of individuals in a similar way as the symptoms of these disorders. Brute-force approaches may learn to exploit effects of these confounding factors on social signals in place of effects due to mental and neuro-developmental disorders. The main objective of this thesis is to develop, investigate, and propose computational methods to screen for mental and neuro-developmental disorders in accordance with descriptions given in the Diagnostic and Statistical Manual (DSM). The DSM manual is a guidebook published by the American Psychiatric Association which offers common language on mental disorders. Our motivation is to alleviate, to an extent, the possibility of machine learning algorithms picking up one of the confounding factors to optimise performance for the dataset – something which we do not find uncommon in research literature. To this end, we introduce three new methods for automated screening for depression from audio/visual recordings, namely: turbulence features, craniofacial movement features, and Fisher Vector based representation of speech spectra. We surmise that psychomotor changes due to depression lead to uniqueness in an individual’s speech pattern which manifest as sudden and erratic changes in speech feature contours. The efficacy of these features is demonstrated as part of our solution to Audio/Visual Emotion Challenge 2017 (AVEC 2017) on Depression severity prediction. We also detail a methodology to quantify specific craniofacial movements, which we hypothesised could be indicative of psychomotor retardation, and hence depression. The efficacy of craniofacial movement features is demonstrated using datasets from the 2014 and 2017 editions of AVEC Depression severity prediction challenges. Finally, using the dataset provided as part of AVEC 2016 Depression classification challenge, we demonstrate that differences between speech of individuals with and without depression can be quantified effectively using the Fisher Vector representation of speech spectra. For our work on automated screening of bipolar disorder, we propose methods to classify individuals with bipolar disorder into states of remission, hypo-mania, and mania. Here, we surmise that like depression, individuals with different levels of mania have certain uniqueness to their social signals. Based on this understanding, we propose the use of turbulence features for audio/visual social signals (i.e. speech and facial expressions). We also propose the use of Fisher Vectors to create a unified representation of speech in terms of prosody, voice quality, and speech spectra. These methods have been proposed as part of our solution to the AVEC 2018 Bipolar disorder challenge. In addition, we find that the task of automated screening for ASD is much more complicated. Here, confounding factors can easily overwhelm socials signals which are affected by ASD. We discuss, in the light of research literature and our experimental analysis, that significant collaborative work is required between computer scientists and clinicians to discern social signals which are robust to common confounding factors.