Bayesian network approaches in mental health: supporting clinical decision-making for screening, prevention and treatment recommendations

Machine learning predictive tools are driving new approaches in the context of mental health. Recent literature highlights the increasing use of data-driven methods in mental health, emphasising their significance and the ongoing need for research in the field. However, these data-driven methods have some limitations as most of them do not address the following three points simultaneously: 1) Presenting outputs in an explainable way to increase trust of clinicians in data-driven solutions; 2) Utilising approaches which are beneficial for addressing uncertainty, especially given the high prevalence of comorbidities in mental health disorders; 3) Providing practical support to help clinicians make informed decisions.

The aim of this thesis is is to support clinical decision-making by providing explainable, probabilistic, data-driven solutions for mental health care.

The methodologies developed focused on three application areas: screening, treatment recommendations, and prevention. Specifically, the analysis included cases studies on depression, panic disorder, social phobia, specific phobia, and anxiety. The approaches primarily utilised Bayesian networks as the core methodology, along with machine learning techniques such as logistic regression, XGBoost, naive bayes, support vector machines and stochastic gradient descent.

In the screening application, a methodology was developed to identify individuals likely to exhibit depressive symptoms for further screening. A Bayesian network was employed to select the features, and a stochastic gradient descent algorithm was applied to develop the model. The output of the model provided a more targeted screening strategy by effectively focusing on individuals at higher risk. Additionally, a tool was developed to help clinicians visualise the benefits of the of the methodology in reducing the number of necessary screening interviews methodology in reducing the number of necessary screening interviews.

For the treatment recommendations application, a methodology was created to select appropriate treatments for four mental disorders: depression, panic disorder, social phobia, and specific phobia. A multiclass classification model was developed using a Bayesian network, combining data-driven approaches with expert knowledge. Expert knowledge played a crucial role in constructing the Bayesian network structure, enhancing explainability. Furthermore, the methodology provided guidelines for clinicians to analyse the model’s output, facilitating decisions about the recommended treatment.

In the prevention application, a methodology was developed to understand the key factors contributing to recurrent anxiety and analyse how these factors could help clinicians prevent it. A novel approach was employed, combining constraint-based methods with bootstrap techniques to identify the best Bayesian network structure, highlighting the most impactful predictors given the data. The methodology also provides an order of importance for these predictors to focus on, aiding in effective prevention strategies.

In all three applications, the methodologies proved suitable for each specific mental disorder case, assisting clinicians in making decisions supported by data-driven analysis. Furthermore, the methodologies have the potential to be extended for use with different mental health disorders.