Approved Research

Longitudinal Sub-Phenotyping of Depression Disorders using Deep Learning Representations

Lay summary

Mental health disorders are a leading cause of burden of disease, affecting over 1 billion people worldwide. In the UK, they are the largest cause of disability, contributing 22.8% of the total burden of disease, ahead of cardiovascular disease (16.2%) and cancer (15.9) at an estimated cost to the economy of £105 billion per year. Among mental health disorders, depression and anxiety disorders have the largest impact.

Depression is an heterogenous disorder, with significant comorbid load, both concurrent psychiatric disorders and other diseases. Depression patients have at least twice the prevalence of nearly all other mental health disorders depending on the severity of presentation. Furthermore, they have higher comorbid prevalence of cervical, thoracic, and intervertebral disc disorders, hypertensive diseases and metabolic disorders, inflammatory and haemorrhagic gastrointestinal diseases, disorders of the urinary and respiratory system, and finally higher utilisation of healthcare resources.

Over the past 20 years our understanding of depression disorders have made tremendous progress, thanks to in depth analysis of clinical presentations, brain morphology and function, and genetic predisposition. Despite this progress, our understanding is still insufficient to explain the disorder heterogeneity, severity of presentation, clinical trajectory and provide precise reasons behind response to treatment. This is perhaps most evident in the fact that current therapies fail to address the increasing need, with only 60% patients with uncomplicated presentation responding to any single antidepressant, and approximately 30% of patients failing to respond to two or more first-line antidepressant, and then experiencing high morbidity, 80% 1-year relapse and 40% 10-year recovery rates.

In this 3-year project we aim to harness UKBiobank via state-of-the-art machine learning methods (deep learning representations) to uncover depression sub-groups (sub-phenotypes) based on patients' temporal clinical trajectory, captured in their electronic health records data as a sequence of diagnostic and treatment actions, and that we hypothesize share a common biological basis. The sub-phenotypes will be fully characterised from both the clinical and genetic perspective, uncovering their clinical, imaging, metabolic, activity (sleep) and genetic traits.

Equipped with a higher resolution understanding we aim to enable the paradigm of precision medicine for depression patients, making possible the delivery of personalised care plans, potentially identifying new drug targets, and providing a foundation for precision clinical trial design. Ultimately, a deeper understanding would enable targeted, timely and efficient resource allocation in the National Health Service, addressing the delivery of care in an environment of increased demand with restricted resources.