Association Between Comorbidity Clusters and Mortality in Patients With Cancer: Predictive Modeling Using Machine Learning Approaches of Data From the United States and Hong Kong.

BACKGROUND

Patients with cancer and cancer survivors often experience multiple chronic health conditions, which can impact symptom burden and treatment outcomes. Despite the high prevalence of multimorbidity, research on cancer prognosis has predominantly focused on cancers in isolation. There is growing interest in machine learning techniques for cancer studies. However, these methods have not been applied in the context of supportive care for patients with cancer who have multimorbidity. Furthermore, few studies have investigated the associations between comorbidity clusters and mortality outcomes.

OBJECTIVE

This study investigated comorbidity clusters among patients with cancer using machine learning and examined their associations with mortality outcomes in two large representative samples from the United States and Hong Kong.

METHODS

This study used data from the National Health and Nutrition Examination Survey (NHANES) and the Hospital Authority Data Collaboration Laboratory (HADCL). Participants aged ≥20 years with a history of cancer were included. The study used a two-step framework to identify clusters of comorbidities in NHANES. In the first step, we used four machine learning techniques, including the Bernoulli mixture model and partition-based methods, to cluster the comorbidities. In the second step, domain experts reviewed and ranked the identified clusters to ensure clinical relevance. The clusters that had the highest average rank were selected for further analysis. The associations between comorbidity clusters and mortality outcomes were analyzed using Cox proportional hazards models. We conducted an external validation to evaluate the generalizability of the clusters identified in the NHANES cohort and their associations with mortality using HADCL. The same number of clusters was replicated based on the distinctive patterns and distribution of comorbidities observed within each cluster.

RESULTS

The study included 4390 participants in NHANES and 12,484 participants in HADCL. Four comorbidity clusters were identified: low comorbidity, metabolic, cardiovascular disease (CVD), and respiratory. In NHANES, participants in the respiratory cluster had the highest risk of all-cause mortality (adjusted hazard ratio [aHR] 1.62, 95% CI 1.26-2.08; P<.001), followed by the CVD cluster (aHR 1.50, 95% CI 1.26-1.80; P<.001) compared to the low comorbidity cluster. The 3 clusters were associated with higher risks of CVD-related mortality (aHR 1.48-3.05, 95% CI 1.14-4.07; P<.003). The effects of comorbidity clusters on mortality were modified by income-to-poverty ratio (P for interaction=.04), diet quality (P for interaction=.02), and cancer prognosis (P for interaction=.005). In the HADCL (validation) cohort, participants in the respiratory and CVD clusters had a higher risk of all-cause mortality.

CONCLUSIONS

High comorbidity burden clusters showed increased all-cause and CVD-related mortality in patients with cancer. These findings highlight the significance of considering comorbidity burden in cancer care. Machine learning approaches can provide valuable insights into complex multimorbidity profiles. Further research is needed to deepen understanding of the relationships between multimorbidity and cancer-specific outcomes.