Mapping drivers of cancer outcomes with holistic analysis using machine learning

For the first time, researchers have used machine learning, a type of artificial intelligence (AI), to identify the most important factors in cancer survival rates in nearly every country in the world.

This study is published in a leading cancer journal Oncology annual report Today (Wednesday) we will provide information on what improvements and policy changes can be made in each country that will have the biggest impact on improving cancer survival rates. The researchers created an online tool where anyone can find their country and scroll down to see which factors, such as national wealth, access to radiotherapy, and universal health insurance, are most associated with cancer outcomes.

“Cancer outcomes around the world vary widely, driven in large part by differences in each country’s health care system,” said study co-lead Dr. Edward Christopher Dee, a radiation oncology resident at Memorial Sloan Kettering (MSK) Cancer Center in New York, USA. “We wanted to create a practical, data-driven framework to help countries identify the most impactful policy tools to reduce cancer mortality and close the capital gap.”

“We found that access to radiotherapy, universal health coverage, and economic empowerment are often important levers associated with improved cancer outcomes nationwide. But other important factors are associated as well.”

Dr. Dee and his colleagues used machine learning to analyze cancer incidence and mortality data from the Global Cancer Observatory (GLOBOCAN 2022) for 185 countries. We also collected information on health systems from the World Health Organization, World Bank, United Nations agencies, and radiation therapy center directories. These include health expenditure as a percentage of GDP, GDP per capita, number of doctors, nurses, midwives and surgeons per 1,000 population, universal health coverage, availability of pathology services, human development indicators, number of radiotherapy centers per 1,000 population, gender inequality index, and proportion of out-of-pocket payments.

Milit Patel created a machine learning model based on data from these global health systems. He is the study’s lead author and a researcher in biochemistry, statistics and data science, and healthcare reform and innovation at the University of Texas at Austin and MSK.

Dr Patel said: “We chose to use machine learning models because they allow us to generate country-specific estimates and associated predictions. Of course, we are aware of the limitations of population-level data, but we hope these findings can help guide global cancer system planning.”

This model produces a mortality-in-morbidity ratio (MIR). MIR reflects the proportion of cancer cases that result in death and serves as a proxy for the effectiveness of cancer treatment. It uses a method that explains individual predictions by quantifying the contribution of each factor to the prediction. This is called SHAP (Shapley Additive exPlanations).

“Beyond simply accounting for disparities, our approach provides policymakers with an actionable, data-driven roadmap to pinpoint which health system investments will have the greatest impact for countries,” said Patel. “This research will help countries prioritize resources and close survival gaps in the most equitable and effective way possible. International organizations, health care providers, and advocates may also use web-based tools to highlight areas for investment, especially in resource-limited settings.”

Using Brazil as an example, the model shows that the factor with the greatest positive impact on mortality versus morbidity is universal health coverage (UHC), while pathology services and nurses and midwives per 1000 population may have a less significant impact on outcomes. Researchers say this suggests that Brazil should prioritize UHC.

As another example, in Poland, the density of radiotherapy services, GDP per capita, and UHC index were shown to have the greatest impact on cancer outcomes, among other key factors. The findings suggest that current efforts to strengthen access to health insurance and services have more pronounced benefits than factors such as general health spending, which have smaller positive effects.

Data show that all health system factors are associated with improved cancer outcomes in Japan, the United States, and the United Kingdom, with the density of radiotherapy centers in Japan and GDP per capita in the United States and United Kingdom having the greatest impact. This suggests that these are the factors that policymakers should focus on.

In China, the situation is more complicated. GDP per capita, UHC, and radiotherapy center density are the factors most contributing to improved cancer outcomes. Out-of-pocket costs, number of surgical personnel per 1000 population, and health spending as a share of GDP are factors that are currently unlikely to explain differences in cancer outcomes.

In their paper, the researchers write about China: “Despite national improvements in health financing and access, high direct costs to patients remain a significant barrier to optimal cancer outcomes. These findings highlight the rapid development of China’s health system. While this has led to important gains in cancer control, gaps in economic protection and coverage persist, highlighting the need for stronger policies focused on reducing out-of-pocket costs and further strengthening UHC implementation to maximize impact on health systems.

Mr. Patel explained the importance of the green and red bars that appear on each country’s graph. “The green bars represent the factors that we currently believe are most strongly positively associated with improved cancer outcomes in a given country. These are areas where we expect continued or increased investment. Deliver meaningful impact. However, red bars do not indicate that these areas are unimportant or should be ignored. Rather, they reflect areas that are currently unlikely to explain the largest differences in outcomes, according to the model and current data. This may be due to already good performance on these aspects, limitations in the available data, or other context-specific factors.

“Importantly, the appearance of a ‘red’ bar should not be interpreted as a reason to discontinue efforts to strengthen pillars of cancer care; improvements in these areas are still of value to a country’s health system as a whole. Our results suggest that if the goal is to maximize improvements in cancer outcomes as defined by the model, focusing first on the strongest positive (green) drivers may be the most impactful strategy.”

The strengths of this study are that it covers nearly all countries, uses the latest global health data, provides actionable country-specific policy recommendations (rather than just global averages), and provides a more explainable AI model. Limitations include: Reliance on aggregated national data rather than individual patient records. Variations in the quality of registries and data, especially in many low-income countries. Domestic trends may overlook within-country disparities, which merit further investigation. And this study cannot prove that focusing on a particular area causes improvements in cancer outcomes, only that such efforts may be associated with cancer outcomes. The findings can help policy makers set priorities, but further research is needed to intervene in specific areas.

Dr Dee concluded: “As the global burden of cancer increases, this model will help countries maximize impact with limited resources. It will turn complex data into understandable and actionable advice for policy makers, enabling accurate public health outcomes.”