The machine learning model calculated cancer mortality and morbidity ratios for each country and assessed the factors that contribute most to differences in survival rates between countries. Additionally, artificial intelligence (AI) tools have mapped out actions that countries can take to improve cancer outcomes. A report describing the establishment of a machine learning framework and country-specific findings is Oncology annual report.
“Cancer outcomes around the world vary widely, primarily due to differences in national health systems. We wanted to create a practical, data-driven framework that would enable countries to identify the most impactful policy tools to reduce cancer mortality and close the capital gap,” said co-principal investigator and co-corresponding author. Dr. Edward Christopher DeeRadiation oncology resident at Memorial Sloan Kettering Cancer Center. “We found that access to radiotherapy, universal health coverage, and economic empowerment were often important levers associated with improved cancer outcomes nationwide. But other important factors were associated as well.”
Background and training method
Cancer survival statistics tend to vary widely from country to country for a variety of reasons, including policies, resources, economic circumstances, and access to healthcare. The researchers aimed to investigate country-specific contributors to global inequalities in cancer treatment and survival outcomes through an explainable machine learning framework.
They collected mortality and morbidity rates for 185 countries from data from the Global Cancer Observatory (GLOBOCAN 2022) and indicators of national health system health from the World Health Organization, World Bank, United Nations, and others. Factors collected that may drive country-specific cancer outcomes include gross domestic product per capita, universal health coverage index, radiotherapy centers per capita, health spending, out-of-pocket spending, workforce density, pathology access, human development index, and gender inequality index. The researchers used the open-source CatBoost library to train and cross-validate the tree-based algorithm, excluding one country from the evaluation during each training session until a total of 1,850 predictions were made. We limited bias and overfitting through nested hyperparameter optimization and applied tight controls to the model to prevent leakage between the training and evaluation sets. SHapley Additive exPlanations (SHAP) assesses country-specific and global characteristics and converts them into values that indicate how much each factor contributes to cancer outcomes.
Overall findings
Strong out-of-sample performance was seen from models with high ability to explain all variability (R2 = 0.852; 95% confidence interval [CI] = 0.801–0.891), the predicted error magnitude is lower (root mean square error = 0.057; 95% CI = 0.050–0.064). The correlation coefficient between predicted mortality and observed mortality-to-incidence ratio was 0.923 (P = 8.30 × 10-78).
At the global level, SHAP values consider gross domestic product per capita (22.5%), radiotherapy centers per population (15.4%), and universal health coverage index (12.9%) to be the most important determinants of cancer outcomes. The most important country-specific drivers varied widely depending on the context, with proposals tailored to infrastructure, coverage expansion, financial protection, and other country-specific recommendations. Consistently cited drivers included radiotherapy capacity and universal health coverage, and the model highlighted the need for more strategic allocation of health spending in many countries, rather than simply increasing funding and resources.
Researchers created an online tool to access country-specific analyzes and take recommended actions to improve cancer outcomes globally.
“Rather than simply explaining disparities, our approach provides policymakers with an actionable, data-driven roadmap to pinpoint which health system investments will have the greatest impact for countries. As the global burden of cancer increases, these insights will help countries “It can help prioritize resources and close survival gaps in the most equitable and effective way possible. International organizations, health care providers, and advocates could also use web-based tools to highlight areas for investment, especially in resource-limited settings,” said the first author. Mirit Patelresearchers at the University of Texas who built the machine learning model.
Country-specific findings
A machine learning model showed that although universal health coverage has had a positive impact on mortality and morbidity in Brazil, this ratio has been negatively affected by reduced access to pathology services and a decline in the proportion of nurses and midwives in the total population.
In the case of Poland, the model suggested that efforts to improve health insurance and increase access to health services had a greater impact on cancer outcomes than general health spending. The biggest influence on improving cancer outcomes in Japan was the density of radiotherapy centers. In both the United States and the United Kingdom, gross domestic product per capita has the greatest impact on mortality and morbidity, suggesting that this is the most impactful area for these countries to focus on for even greater growth.
“As the global burden of cancer increases, this model can help countries maximize impact with limited resources. It turns complex data into understandable and actionable advice for policymakers, enabling accurate public health care,” Dr. Dee concluded.
Disclosure: This research was partially funded through a Prostate Cancer Foundation Young Investigator Award and a Cancer Center Support Grant from the National Cancer Institute. For full study author disclosures, please visit annalsofoncology.org.
