Machine learning maps the slide into multimorbidity

Hypertension, diabetes, angina, stroke and obesity rarely stay solitary, and once two or more coexist — cardiometabolic multimorbidity — care becomes more complicated and outcomes worse. Prevention would be far easier if it were possible to identify who is likely to move from healthy to a first disease, and separately who is likely to move from one disease to several. Most risk scores treat disease as a single yes/no endpoint rather than a journey through stages. This study asked whether a broad set of personal, behavioural, economic and environmental factors could predict each step of that journey in middle-aged and older Chinese adults.

The researchers drew on 4,518 participants from the World Health Organization's Study on Global AGEing and Adult Health in China, covering the period 2007 to 2019. Cardiometabolic disease incidence was captured through self-reported surveys, and multimorbidity was defined as the presence of at least two of hypertension, diabetes, angina, stroke and obesity. A multi-state model examined which multidimensional factors influenced the transition from health to a single disease, and then from a single disease to multimorbidity. The retained predictors were fed into a stacking ensemble model combining five machine-learning algorithms, with discrimination reported as area under the curve.

During follow-up, 52.19% of initially healthy participants developed one cardiometabolic disease, and among those, 15.61% progressed to multimorbidity. Female sex, low GDP per capita, unhealthy behaviours, elevated PM2.5 concentrations, low humidity and low temperatures were shared risk factors across both transitions. Older age, low educational level and physical fitness impairment were independent risk factors only for the move from health to a first disease, whereas impaired intrinsic capacity was distinctive to the progression from one disease to multimorbidity. The ensemble model reached an AUC of 0.89 (95% CI 0.88-0.91) for the health-to-disease transition and 0.76 (95% CI 0.72-0.82) for disease-to-multimorbidity.

The finding that behaviour, air pollution and socioeconomic position all track cardiometabolic risk is consistent with a large existing literature, but this analysis adds the idea that the risk profile is not identical at every stage. Physical fitness impairment mattered for getting the first diagnosis, while intrinsic capacity mattered for accumulating a second — a staged picture that complicates one-size-fits-all risk scoring. Because disease was self-reported and the model was built and tested in a single national cohort, the AUC figures are optimistic estimates of how such a tool would work in practice. External validation in other populations, and prospective testing of whether targeting the identified factors actually slows progression, remain open.

This is a reason to think of cardiometabolic disease as a sequence rather than a single event, with different levers at different points along it. Factors you can partly influence — behaviour and physical fitness — showed up alongside factors you largely cannot, such as air pollution, climate and local economic conditions, which is a reminder that risk is environmental as well as personal. The reported accuracy figures describe a research model, not a tool available to a reader today. Nothing here identifies an intervention that has been shown to change the trajectory; it maps where risk concentrates.