Can machine learning be used to predict antidepressant use outcome longitudinally?

点击翻译按钮获取中文摘要

This project aims to understand the long-term effects of antidepressant use in the adult population. There is currently a gap in the literature regarding the longitudinal effects of pharmaceutical interventions for depression, so there is not great clinical understanding of the impact of antidepressants on health outcomes long-term. For example, some longitudinal studies suggest that the symptoms of depression improve overall after long term antidepressant use, but this is reported in tandem with adverse psychological and physical health outcomes like weight gain, sexual dysfunction, and emotional numbness (Dehar et al. 2016). In the psychiatric literature, there is an impoverished understanding of depression causality, with multiple competing hypotheses suggesting genetic, psychological, neurochemical, and neurostructural correlates of the illness. This suggests that depression as an overarching umbrella term could include multiple phenotypes, that if captured, could explain different illness trajectories and predict differential health outcomes after pharmaceutical intervention.The long-term effects of antidepressant use on a wide range of health outcomes should be explored to ascertain whether they are an optimal intervention for people who present with mild to moderate depression and anxiety. This project will apply machine learning and modelling techniques to the Avon Longitudinal Study of Parents and Children (ALSPAC) data set, to develop predictive models of antidepressant use outcomes. For example, significant changes in patient symptom profiles may indicate cognitive, psychological, or behavioural changes after long-term antidepressant use.Initial exploration of the ALSPAC data set will be undertaken using a standard data science workflow, to analyse and understand exposures relevant to antidepressant use outcomes. Advanced machine learning methods will be used, and potentially progressed, in order to optimise novelty of contribution. State of the art anomaly detection will be developed to identify exposures that significantly change outcome probability. Bayesian network models and causal learning methods will be developed to compute the probability of relevant exposures being causally related to measurable patient outcomes, such as behavioural, cognitive, psychological and lifestyle outcomes. Deep clustering algorithms will be used to sub-group patients together, to explore whether certain exposures have more predictive power in long-term mood outcomes than others and factors which may modify associations such as genetic profiles or engagement with psychological therapies .The timeline of the project is still being planned, but the first year would include developing models and running them on test datasets, bespoke dataset construction from existing ALSPAC data, and ALSPAC data familiarity. The second year would be used to run iterations of optimised models and potentially replicate with other cohort studies, with the final year being used for thesis write up and communication of results to relevant clinical bodies such as psychiatrists, GPS, and policy makers through digital health / machine learning publications.The novelty of the project lies in the application of machine learning techniques to ALSPAC data to track antidepressant outcomes over at least a decade. While ML techniques have been applied across a number of digital healthcare domains, their use for mental health monitoring, prediction, and understanding the clinical pathways of patients remains under-researched. This is critically important moving forward, as this work will feed into the increasing use of precision medicine and apps for mental health (e.g., digital phenotyping), whereby having a deeper understanding of the phenotypes underpinning depression and how we can analyse this quickly and efficiently could transform the way we intervene and offer healthcare.