Machine learning tools predict economic harm to cancer patients

Researchers at MUSC Hollings Cancer Center have developed a machine learning tool to identify cancer patients at high risk for financial toxicity, the financial stress and challenges associated with cancer diagnosis and treatment. The study brought together several researchers from the Hollings Cancer Prevention and Control Research Program and reflected the project’s focus on cancer outcomes, survival, and care delivery.

The study, published in JNCI Cancer Spectrum, describes a personalized risk prediction model that uses patient information to estimate the likelihood that someone will suffer from cancer-related financial burdens, such as medical debt, unpaid bills, and fear of treatment costs.

Unfortunately, cancer treatment is expensive and the economic fallout is a complex issue. You will incur travel costs, lodging costs, loss of income, medical expenses, and out-of-pocket costs. In some cases, financial stress causes patients to delay or even discontinue treatment. We aimed to identify people at risk early, before these challenges escalate. ”

Dr. Haluk Damgacioglu, first author

Connect patients to support faster

Nearly one-quarter of cancer patients in the United States experience economic toxicity, which includes both financial hardship and psychological stress.

“There is a material aspect, such as debts and unpaid bills, but there is also a psychological aspect,” Damgacioglu explained. “Just worrying about how to pay for treatment can be a huge source of stress.”

Although much research has been conducted on who is most likely to experience financial hardship during cancer treatment, methods to predict patient risk remain limited. Early identification may connect at-risk patients to counseling and other services before financial burden impacts care decisions, treatment adherence, and quality of life.

Hollings provides a wide range of patient services to patients and their families. This includes financial counseling from a counselor who specializes in cancer treatment.

“Hollings has resources for financial navigation and counseling,” Damgasioglu said. “The first step is to identify patients who may need additional support so we can provide those resources sooner.”

Building tools to predict financial risk

To address this gap, the research team analyzed national survey data from nearly 800 cancer patients who were undergoing or had completed cancer treatment within the past year. Patients were classified as experiencing financial toxicity if they answered “yes” to at least one of several material hardship or psychological stress questions, including borrowing money, being unable to pay medical bills, having debt, filing for bankruptcy, or worrying about future medical costs related to cancer treatment.

Researchers tested six machine learning models that used patient demographic, clinical, and financial information to predict who would experience financial toxicity. They then fine-tuned the model to maximize sensitivity and prioritize identifying as many at-risk patients as possible.

“We didn’t want to miss out on anyone who might experience economic harm,” Damgasiolu said. “That was one of the most important goals of the study.”

The best-performing model identified patients at risk for economic toxicity with 84% sensitivity and 75% specificity, balancing the ability to detect patients in need of support with the ability to minimize false alarms. This model identified most patients likely to experience economic toxicity without unnecessarily flagging many false positives.

The study also used interpretable machine learning techniques to identify the factors most strongly associated with financial risk. Some of the strongest predictors include:

• young in age.
• I have little income.
• Your overall health deteriorates.
• Aggressive cancer treatment.
• The out-of-pocket amount of medical expenses will increase.

To translate this research into clinical care, the team developed a publicly available web-based risk calculator that estimates the likelihood that a patient will experience economic toxicity and categorizes it as low, moderate, or high. The researchers envision this tool eventually helping connect patients with financial and other support services early in their treatment. Their future work will focus on refining the model, testing it in real-world clinical settings, and investigating how financial stress affects patients’ long-term health outcomes.

“Economic toxicity is also a side effect of cancer,” Damgasioglu said. “If we can identify risk early, we may have an opportunity to help patients before stress becomes too great.”