We analysed a nationally representative population of 5,846,480 million people in the UK, grouping their individual trajectories based on the chronological order of disease development. A key innovation of our method is the use of an LLM to cluster patients rather than diseases, leveraging the full spectrum of clinical information – including prescriptions, diagnoses, and laboratory tests – representing a major advance on earlier work that has primarily mapped disease co-occurrence cross-sectionally. By employing an advanced transformer architecture (EHR-DeBERTa), our approach facilitates a dynamic and more personalised understanding of MLTC patterns, resulting in interpretable gender-specific patient clusters. This approach has significant applications in identifying shared disease pathways, enhancing risk stratification, and optimizing healthcare delivery and management strategies.

We identified five main groups of clusters in both female and male patients: clusters of relatively young and healthy individuals, clusters of middle-aged individuals dominated by mental health conditions, clusters of older individuals with cardiometabolic comorbidities, clusters of patients with respiratory diseases and comorbidities and clusters of patients with mixed conditions spanning different body systems across a wider age range. Within each broad group, there were different cluster subtypes representing distinct patient trajectories including recognisable co-occurrence of linked conditions as well as other less well studied disease combinations. These findings highlight how MLTC manifests across age groups, for example, with mental health conditions being particularly prevalent in middle age and cardiometabolic diseases become more dominant in older populations. This age-related variation highlights the need for tailored healthcare strategies that address the evolving burden of MLTCs over the life course and emphasizes the importance of early interventions before disease has occurred, for example, to prevent transitions from a low disease burden cluster to a mental health cluster in midlife or a cardiometabolic cluster later in life.

In females, mental health clusters were more frequent, accounting for 28% of females versus only 13% of males, reflecting the higher prevalence and risk of mental health – particularly depression and anxiety – in females compared to males.²⁹ In both genders, we identified a highly comorbid cluster with multiple mental health conditions, co-occurring with diabetes, obesity, and neurological conditions including Parkinson’s Disease and multiple sclerosis.^30,31 These findings reinforce existing epidemiological evidence on bidirectional associations between these conditions³² and further stresses the need for integrated management approaches that address both mental and physical health within a unified care framework.

We observed significant variation in the prevalence of cardiometabolic conditions across clusters, suggesting that these conditions play a dominant role in distinguishing patient phenotypes. Cardiometabolic clusters were more prevalent in males compared to females and females in cardiometabolic clusters tended to be older than their male counterparts, consistent with literature.^33,34 However, the composition of clusters showed notable similarities between genders. In both males and females, we observed a highly comorbid cluster of cardiovascular, metabolic and renal comorbidities, often co-occurring with other linked conditions such as gout, COPD, and neurodegeneration. This underscores the well-recognised bidirectional association between the dysfunction of the heart, kidneys and metabolic system, and their systemic multi-organ consequences.³⁵ The co-occurrence of neurodegeneration highlights an important heart-brain-liver axis which is less well-characterised. As more therapeutic options emerge, it is crucial to develop integrated treatment strategies that maximize multi-organ benefits and establish shared guidelines for managing these interconnected conditions. Simultaneously, a multi-organ approach to disease prevention could enable to early intervention as abnormalities in any of these organs may serve as early indicators of disease progression, offering opportunities for timely intervention, and improved patient outcomes. Interestingly, mental health and cardiovascular clusters were mutually exclusive: clusters with a high prevalence of cardiometabolic conditions had a lower prevalence of mental health conditions and vice versa. This contrasts with strong epidemiological evidence of increased incidence of cardiovascular disease in those with pre-existing mental health conditions, particularly psychotic disorders.³⁶ Possible explanations include the younger age of those in the mental health clusters, who may not yet have developed cardiovascular disease, and the older age of those in the cardiometabolic clusters, mental health conditions may be underdiagnosed due to stigma and barriers to accessing support.³⁷

Other multi-organ effects were evident from the patient clustering. For example, the clustering of patients with IBD with COPD and other respiratory conditions. These conditions may share similar underlying pathways such as infection, immune activation, inflammation, and shared microbiota effects which may lead to common treatments and monitoring of patients with IBD for the potential of developing respiratory disease.³⁸ Similarly, the association between osteoporosis and dementia raises the possibility of an underlying biologic mechanism linking bone loss and cognitive decline, especially in females.³⁹

The identified clusters of patients differed substantially based on deprivation and ethnic diversity. Within each broader patient group, individuals with similar ethnic backgrounds and socioeconomic status tended to cluster together. For example, we identified distinct male cardiometabolic clusters with an overrepresentation of White individuals (CM2), South Asian individuals (CM1), and Black individuals (CM6), with the latter two clusters having a lower median age than CM2. We also found that low disease burden clusters were underrepresented by people living in more deprived areas. Several conclusions can be drawn. Firstly, MLTCs are a widespread issue across the population and although not confined to populations with specific ethnic or socioeconomic backgrounds, people in low disease burden clusters are more likely to be from less deprived areas. Secondly, individuals with similar socioeconomic status and ethnic backgrounds may experience comparable trajectories of comorbidities, reflecting shared risk factors within these groups. Thirdly, clusters including people living in areas of greater deprivation and from non-White ethnic backgrounds tend to be younger, possibly indicating earlier onset of MLTCs in these subgroups. Previous research has also highlighted that people living in more deprived areas have a faster accumulation of conditions and higher mortality rates.⁴⁰ Collectively, our findings emphasise that tailored healthcare strategies are essential to address the diverse needs of different population groups in addition to integrated approaches to managing MLTC.

Traditional clustering approaches in healthcare often focus on either patient characteristics or clinical outcomes.¹⁹ LLMs offer an opportunity to combine both, by fine-tuning on outcomes such as mortality or hospital admission. This could enable the generation of clusters that not only reflect patient characteristics but also predict downstream risks. Future research could investigate the relationship between model self-attention and cluster assignment to identify sequences of diseases which are at higher risk of adverse health outcomes. Simulated interventions can also be explored by modifying patient histories (e.g., adding medication at a specific time point or delaying a diagnosis), informing personalised lifestyle, or treatment recommendations and drug repurposing. Fine-tuned models could also support clinical decision-making by providing the probability of a missed diagnosis or the potential effect of an interventions for each patient. Another strength of this study is the use of a large, contemporary dataset which is representative of the English population.²⁰ An advantage of primary care data is that historic diagnoses can be entered retrospectively with the date of diagnosis, capturing a more complete view of a person’s disease history than is often available in hospital records.

Some limitations also need to be acknowledged. Although the prevalence of many common diseases in CPRD is similar to other national data sources,²¹ for conditions such as cancer, diagnosis recording is less complete than estimates derived from gold standard sources such as cancer registry data.²² The presence or absence of a diagnostic code in GP EHR data is also influenced by factors independent of a person’s health, such as GP practice organisational coding policies and financial incentives, potentially biasing model learning towards conditions which are recorded more frequently.⁴¹ We summarised continuous diagnostic tests as binary outcomes (e.g., abnormal C-reactive protein levels) which may result in loss of information. Future work could explore use of additional embedding layers that can provide information on continuous variables, such as blood pressure readings. We utilised a hard clustering method, using a robust method to ensure the optimal cluster number and evaluated the quality of the clustering with multiple methods. However, the optimal number of clusters is difficult to define as there may be multiple optima with similar performance. Future studies employing soft-clustering methods may allow for more flexible identification of patient clusters. We presented the single diseases that were prevalent within each cluster. While we presented individual diseases prevalent in each cluster, the underlying embeddings also captured disease sequences over time. Future work could explore whether common diseases trajectories differ between clusters that otherwise appear similar.

We demonstrated the potential of an LLM-based clustering pipeline to identify gender-specific clusters of patients and to provide a more nuanced and personalised understanding of MLTC patterns by leveraging longitudinal EHR data. Our findings also indicate valuable opportunities to explore shared disease mechanisms and potential drug repurposing applications. Cardiometabolic and mental health conditions emerged as the dominant drivers of distinct patient phenotypes, with notable differences in males and females. We observed clear differences in MLTC composition across the lifespan, with mental health conditions more dominant in middle age and cardiovascular diseases more prevalent in older adults, and with unique MLTC clusters shaped by differences in ethnic and socioeconomic backgrounds. Collectively, these findings underscore the need for integrated health services and interventions, which are patient-centred, rather than disease-centred. Future work should focus on refining clustering methodologies, incorporating clinical outcomes, and exploring the utility of clusters to inform real-world models of care in healthcare settings. Ultimately, integrating LLM-driven patient clustering into clinical practice could enhance risk stratification, enable early detection of conditions and improve long-term health outcomes for people with MLTC.