A machine-learning model has calculated country-specific cancer mortality-to-incidence ratios and evaluated the factors that contribute the most to each country's survival gaps. Additionally, the artificial intelligence (AI) tool mapped out actions each country could take to improve cancer outcomes. A report explaining the establishment of the machine learning framework and its country-specific findings was published in Annals of Oncology.
“Global cancer outcomes vary greatly, largely due to differences in national health systems. We wanted to create an actionable, data-driven framework that helps countries identify their most impactful policy levers to reduce cancer mortality and close equity gaps,” said co-lead researcher and co-corresponding author Edward Christopher Dee, MD, Resident Physician in Radiation Oncology at Memorial Sloan Kettering Cancer Center. “We found that access to radiotherapy, universal health coverage, and economic strength were often important levers being associated with better national cancer outcomes. However, other key factors were relevant as well.”
Background and Training Methods
Cancer survival statistics tend to vary significantly country to country for many reasons, including policy, resources, economic health, health-care access, and more. Researchers aimed to explore country-specific contributors to global inequities in cancer care and survival outcomes through an explainable machine-learning framework.
They gathered mortality-to-incidence ratios for 185 countries from Global Cancer Observatory (GLOBOCAN 2022) data, as well as indicators of the health of the country's health-care system from the World Health Organization, the World Bank, the United Nations, and more. Factors gathered that could drive country-specific cancer outcomes included: gross domestic product per capita, universal health coverage index, radiotherapy centers per population, health spending, out-of-pocket expenditure, workforce densities, pathology access, Human Development Index, and gender inequality index. The research team trained a tree-based algorithm using the open-source CatBoost library to cross-validate, with one country left out for evaluation during each training until a total of 1,850 predictions were made. They limited bias and overfitting with nested hyperparameter optimization and applied strict controls to the model to prevent leakage between training and evaluation sets. With SHapley Additive exPlanations (SHAP), country-specific and global features were evaluated and translated into values for how much each factor contributed to cancer outcomes.
Overall Findings
A strong out-of-sample performance was seen from the model with a high ability to explain all variability (R2 = 0.852; 95% confidence interval [CI] = 0.801–0.891) and a low predicted magnitude of error (root mean squared error = 0.057; 95% CI = 0.050–0.064). The correlation coefficient for between the predicted and observed mortality-to-incidence ratios was 0.923 (P = 8.30 x 10-78).
On a global level, the SHAP values considered gross domestic product per capita (22.5%), radiotherapy centers per population (15.4%), and universal health coverage index (12.9%) to be the most significant determinants of cancer outcomes. The most significant country-specific drivers varied widely by settings, leading to tailored suggestions for infrastructure, coverage expansion, financial protection, and other recommendations on a country-by-country basis. Drivers that were consistently raised included radiotherapy capacity and universal health coverage, and the model highlighted that more strategic allocation of health spending is needed in many countries beyond just increased funds and resources.
The researchers created an online tool for accessing the country-specific analyses and recommended actions to improve cancer outcomes globally.
“Beyond simply describing disparities, our approach provides actionable, data-driven roadmaps for policymakers, showing precisely which health system investments are associated with the greatest impact for each country. As the global cancer burden grows, these insights can help nations prioritize resources and close survival gaps in the most equitable and effective way possible. International organizations, health-care providers, and advocates may also use the web-based tool to highlight areas for investment, especially in resource-limited settings,” said lead author Milit Patel, a researcher at the University of Texas, who built the machine-learning model.
Country-Specific Findings
The machine-learning model showed that the universal health coverage has positively impacted the mortality-to-incidence ratio for Brazil, but that the ratio was hurt by lower access to pathology services and a lower rate of nurses and midwives to the overall population.
For Poland, the model suggested that efforts to improve health insurance and access to more health services has greatly impacted cancer outcomes—more so than general health spending. The density of radiotherapy centers impacted improved cancer outcomes the most for Japan. Gross domestic product per capita had the greatest effect on mortality-to-incidence ratios for both the United States and the United Kingdom, suggesting that this was the most impactful area to focus on for even greater gains going forward for these countries.
“As the global cancer burden grows, this model helps countries maximize impact with limited resources. It turns complex data into understandable, actionable advice for policymakers, making precision public health possible,” Dr. Dee concluded.
DISCLOSURE: The study was funded in part through the Prostate Cancer Foundation Young Investigator Award and through the Cancer Center Support Grant from the National Cancer Institute. For full disclosures of the study authors, visit annalsofoncology.org.
