T41. IDENTIFYING SYMPTOM-BASED SUBTYPES OF DEPRESSION USING MACHINE LEARNING IN THE UK BIOBANK

Major depression is diagnosed based on the presence of symptoms associated with the disorder. However, this allows individuals to present with markedly different symptom profiles that may reflect distinct aetiologies. Identifying common patterns of symptoms using data-driven approaches could help clarify the heterogeneity of depression. Furthermore, examining the demographic characteristics, health status, and polygenic scores of individuals with specific symptom profiles may offer insights into underlying causes. The UK Biobank’s Mental Health Questionnaire (MHQ) released in 2017, and the Mental Well-being Questionnaire (MWQ) released in 2023 were used to conduct this research. Four non-overlapping sets of individuals were identified as “ever depressed” (MHQ n = 17,435 and MWQ n = 14,301) or “recently depressed” (MHQ n = 3,096 and MWQ n = 3,240). Two machine learning approaches, (1) multivariate Bernoulli-mixtures and (2) agglomerative hierarchical clustering, were used to identify common sets of symptoms and cluster individuals by symptom similarity. Replication of results were examined across the two questionnaires and between clustering methods. Regression models were used to examine associations between a cluster and demographic variables (sex, age, body mass index, smoking status, ethnicity, and deprivation), eight health conditions (including stroke, heart attack, and dementia), and polygenic scores for bipolar disorder, schizophrenia, and attention-deficit/hyperactivity disorder (ADHD). Twelve clusters of individuals were identified for “ever depressed” in the MHQ and the MWQ, based on common sets of symptoms. For “recently depressed”, there were 11 clusters of individuals identified in both the questionnaires. There were very strong correlations (r > 0.81) between cluster symptoms across the two questionnaires for “ever depressed” and “recently depressed”. Individuals were also found to be cluster similarly when comparing the two machine learning models (Rand Index > 0.83). The most prevalent cluster in each of the four sets of non-overlapping individuals contained endorsement of all depressive symptoms. Individuals with this symptom profile were younger (P < 1.40 × 10-6), more likely to be female (P < 0.004), have a higher body mass index (P < 2.17 × 10-9) and a higher polygenic score for ADHD (P < 1.14 × 10-5). Machine learning is an underutilised tool in depression research. Clustering algorithms were applied to symptom data and demonstrated replicable results for identifying sets of common symptoms and for the grouping of individuals based on those symptoms. Within these groups individuals shared key demographic variables, health traits, and genetic features. The identification of symptom-based subtypes highlights opportunities for risk stratification with potential implications for improving diagnosis, prognosis, and treatment. This research emphasises the need for continued research into the diverse manifestations of depression.