Identifying Clinical Sub-Groups Using Machine Learning: Functional Status, Symptom Burden, and Healthcare Utilisation...

BackgroundAs the patterns of chronic complex disease and death continue to evolve, the demand for palliative care continues to increase. Understanding key patterns in the population using machine learning allows for new insights.AimUsing a large prospectively collected dataset on function and symptom burden to explore clinically distinct sub-groups of palliative care. These sub-groups were used to explore patterns of healthcare utilisation.DesignThe Australian national Palliative Care Outcomes Collaboration data were used for K-means clustering and multivariable regression.Setting/participantsAdults aged ≥18 years who received palliative care and died between 1 January 2014 and 31 December 2023 (261 290 patients; 350 512 episodes of care).ResultsThree stable episode-level clinical sub-groups were identified: • Sub-group 1 (119 901 episodes; 42.7%): dependency-led; very poor function; predominantly inpatient; shortest episode duration. • Sub-group 2 (57 453 episodes; 20.5%): symptom-led; intermediate function; community and inpatient care; mid-length duration. • Sub-group 3 (103 179 episodes; 36.8%): function-preserved; mild symptoms; community-based care; longest episode duration. In regression, Sub-group 1 and Sub-group 2 had exponentiated coefficients of 0.49 (95% CI 0.49-0.50) and 0.71 (0.70-0.72), respectively, relative to Sub-group 3. Between-cluster differences were greater among non-cancer episodes and varied according to proximity to death; these associations were attenuated after excluding episodes ending in death. Held-out R² was 0.261 for the cluster-based model and 0.316 for the individual-clinical-variable model.ConclusionsK-means clustering provided a concise, clinically interpretable representation of heterogeneity in specialist palliative care. Episode-level clinical sub-groups captured meaningful variation in service use while retaining much of the information contained in clinical measures, supporting prospective evaluation for service planning and care delivery.