BERTopic-based temporal analysis to discern chronic disease multimorbidity patterns

Unhealthy lifestyles and aging demographics contribute to an increased prevalence of multimorbidity, which significantly affects health outcomes by increasing demand for health and social services. Therefore, public health policies and healthcare services must go beyond a singular disease-focused approach and develop a broader perspective of multiple diseases. We used sample cohort data of the Korean National Health Insurance Service from 2002 to 2019 for individuals aged ≥ 40 years with at least two chronic diseases, each with an incidence ≥ 1%, and a disease history spanning at least six years. This study employed BERTopic, a transformer-based deep learning topic modeling framework, to identify long-term multimorbidity patterns and explore interrelationships among chronic diseases. Among 168,529 individuals (73,157 males and 95,372 females), BERTopic identified diverse comorbidity patterns revealing both well-known and previously underrecognized disease associations. In men, cardiovascular and metabolic disease, including ischemic heart disease, diabetes, and stroke, accounted for most patients (20.1%), followed by diabetes associated clusters with musculoskeletal diseases (11.3%). In women, thyroid cancer was associated with musculoskeletal and digestive system disorders (17.5%). This deep-learning-based study of multimorbidity patterns elucidated long-term disease combinations that may result in comorbidities. These findings have significant implications for the prevention and management of chronic diseases and may inform the development of disease pattern-based personalized prevention and intervention methods. The integration of deep learning and medical expertise demonstrates a significant potential for illness forecasting and enhancing patient care efficacy.