Addressing Cancer Mortality Disparities in Appalachia: The Role of AI and Machine Learning in Unraveling SDOH

Patients in Appalachia have become pessimistic about their cancer mortality risk because of the exceptionally high cancer mortality rates in the region compared with the rest of the United States, explained Aisha Montgomery, MD, MPH, senior research manager at Vibrent Health, during a session at the Association of Cancer Care Centers (ACCC) Annual Meeting & Cancer Center Business Summit (AMCCBS) in Washington DC.

Image credit: Peshkova | stock.adobe.com

To address the disparities in minority health and behavioral health in this region, Montgomery and her colleagues launched the National Institutes of Health’s Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD) program, which looks to employ artificial intelligence (AI) and machine learning (ML) to analyze social determinants of health (SDOH) in the underserved populations of rural Appalachia.

“Cancer death rates have decreased by 27% in the US overall, but for some reason, they continue to rise in Appalachia,” Montgomery said. “A caveat to that is that 30 years ago, Appalachia had the lowest cancer mortality rates, but now this has reversed. In rural Appalachia, which is Central Appalachia, we see that the cancer mortality rates are 30% higher than the overall US rates, and even 15% higher than in metro counties surrounding [it].”

According to Montgomery, another unique aspect of cancer rates in Appalachia is specifically high rates of cervical and colorectal cancer, as well as lung cancer because of the environmental conditions of the region. Although cervical and colorectal cancer are known to be effectively prevented through screenings, there are high rates of these cancers occurring in Appalachia.

“We know that there are a lot of social determinants that contribute to cancer mortality and health disparities. In this region, we know that there are transportation issues, there are different levels of health literacy, and there's higher poverty levels and uninsured rates,” Montgomery said. “We wanted to investigate and see how those factors play into the cancer mortality disparity that we're seeing.”

Montgomery noted that during her work in the AIM AHEAD program, she had the privilege of working with undergraduate students. One of the things these students shared with Montgomery is that in rural Appalachia where these students reside, there's a culture present in which the residents feel a sense of fatalism around the potential of getting and dying from cancer. According to these students, residents of the region feel like this fate is inevitable.

“So that's another social determinant that's unique to this area that needs to be addressed,” Montgomery said.

In her work with the AIM AHEAD program, Montgomery conducted a study to analyze SDOH using AI and ML in rural Appalachia in 2 phases.

“We started with a pilot phase,” Montgomery said. “Our study was designed to evaluate the levels of bias in the data. We looked at data from SEER [Surveillance, Epidemiology, and End Results], which is a national cancer registry, and we compared [these data] with the data that was extracted directly from electronic health records [EHRs] at St. Elizabeth Healthcare.”

Additionally, in this pilot phase, Montgomery’s team worked on developing an ML model that could assess whether there was a difference in the mortality prediction for patients with colorectal cancer who live in Appalachia compared with patients who live outside of Appalachia.

“What we saw in our pilot was that the model predicted lower survival in Appalachian patients, so it was biased against [these] patients, and we also saw that the model was less accurate,” Montgomery said. “If we gave it a bunch of patients, it was less likely to identify the patients who actually died if they were from Appalachia.”

In phase 2 of the study, Montgomery’s team decided to see how many SDOH factors are actually in EHR datasets.

“We stratified those to identify the features that had the greatest impact on the prediction of mortality, and we also hypothesized that if we used data from the EHRs vs the national data, we would have a less biased prediction,” Montgomery said. “Our study population for phase 2 was adults who were diagnosed with colorectal cancer from 2000 to 2017. The reason we cut it off in 2017 is because we wanted to try to capture patients who had met that 5-year mark for survival.”

The results of the analysis showed that age had the highest impact on the model’s predictability, which Montgomery noted she does not consider to necessarily be an SDOH. However, after age was marital status, insurance rate, and being a rural patient.

“Those were our top 3 social determinant factors that are affecting survival in our model. Then what we did was we created a model that combines clinical features and social determinant features,” Montgomery said.

In regard to acknowledging marital status as an SDOH, Montgomery noted that could be a “sticky point.”

“But that's what we're seeing so far in the model,” Montgomery said. “Then if you go right below tumor grades, which was our third top clinical feature, you'll see again, that being rural has an effect on cancer outcome. So, these are the things that we're finding so far in our work. Knowing that there is bias in our datasets, that's an issue that we need to address.”

Montgomery noted that one thing her team plans to do is to get more data to expand the regions being assessed and open up the analysis to other rural regions that are affected by high cancer mortality.

“We're partnering with more and more centers so that we can expand the project to the entire southeastern United States, because we see a similar disparity there with morality. We're also trying to foster an infrastructure of sharing to contribute [data],” Montgomery said. “It's very difficult for these small community cancer centers to participate in research, and they don't have the support that they need to share the data because we use deidentified data. So that's something else that we are fostering, because we realize that we can't properly use these advanced technologies if we don't have the right people in our datasets to represent the people who are most affected.”

Reference

Montgomery A. Deep Dive 2. Research and Clinical Trials. Presented at: ACCC AMCCBS; February 28-March 1, 2024; Washington, DC.