Machine learning-based optimisation of ultra-brief questionnaires for mental health disorders

Machine learning (ML) is capable of identifying patterns in vast amounts of data that humans cannot identify as easily. Given this capability, researchers have begun investigating the use of ML in the field of mental health to assist decision-making. This thesis explored the development and application of an ML-based framework to derive the optimal ultra-brief versions of screening questionnaires for depression, generalised anxiety, perceived stress, irrational procrastination, insomnia, and poor general mental health. ML offers a data-driven approach to identifying the optimal symptom combinations at screening for mental health disorders by modelling the underlying function between the disorder symptoms and the diagnosis.

The empirical evidence obtained in this thesis does not support the consensus that the cardinal symptoms of depression or generalised anxiety are the most effective at screening for their presence. The findings also demonstrated the advantages of inputting ultra-brief questionnaire items into ML models compared to using the traditional sum score approach. This thesis contributed to the field of mental health research by introducing novel ML-based ultra-brief depression, generalised anxiety, and insomnia screening instruments that are more effective than the established ultra?brief instruments. Additionally, new 2-item versions of perceived stress, irrational procrastination and general mental health screening questionnaires were developed, filling a notable gap in existing research.

The findings also demonstrated that using a composite 4-item ML-based screening instrument for both depression and generalised anxiety was more effective than using 2-item disorder-specific screening instruments separately. This performance was not replicable by the established depression and generalised anxiety composite screening instrument. Finally, most ML research in mental health has focused on the binary classification of a disorder i.e., depressed or not depressed. Utilising other ML techniques such as regression with ultra-brief questionnaires provided more nuanced predictions during screening regarding the severity of a disorder.