Machine learning algorithms utilizing electronic health records can effectively predict two-year dementia risk among American Indian/Alaska Native adults aged 65 years and older, according to a University of California, Irvine-led study. The findings provide a valuable framework for other healthcare systems, particularly those serving resource-limited populations.
The computer modeling results also found several new predictors for dementia diagnosis that were identified consistently across different machine-learning models. Findings are published in the Lancet Regional Health – Americas. The National Institutes of Health supported the research.
Up until now, no other study has looked at harnessing the power of machine learning models to help predict dementia risk among the historically understudied American Indian/Alaska Native population, as defined by the U.S. Census Bureau.
Machine learning models, which enable computers to make predictions or decisions using vast datasets without explicit programming for each task, enhance efficiency, accuracy and scalability in analyzing large datasets.
The population of older American Indian and Alaska Native adults is projected to increase nearly three-fold between 2020 and 2060. With dementia being a leading cause of disability and mortality in this age group, this debilitating condition is an increasing concern in this community.
In addition to numerous ailments like cognitive decline, weakened immune system and depression, dementia has far-reaching societal impacts. It takes a toll on family members emotionally, incurs substantial medical expenses and contributes to a general decline in quality of life.
“Public health researchers play a significant role in helping clinicians and policymakers make informed decisions about population health,” said Luohua Jiang, a professor of epidemiology and biostatistics at the UC Irvine Joe C. Wen School of Population & Public Health. “If future studies confirm these results, our findings could prove valuable to the Indian Health Service and Tribal health clinicians in identifying high-risk individuals, facilitating timely interventions and improving care coordination.”
Jiang and colleagues took seven years of data from the Indian Health Service’s National Data Warehouse and related electronic health records databases and divided the data into a five-year baseline period (2007 to 2011) and a two-year dementia prediction period (2012 to 2013). The study included nearly 17,400 American Indian/Alaska Native adults aged 65 years or older who were dementia-free at the baseline, of whom almost 60 percent were female.
Over the two-year follow-up, 611 individuals (3.5 percent) were diagnosed with dementia. Four machine-learning algorithms were evaluated and compared based on their data preprocessing efforts and model performance. Of the three top-performing models the team developed, 12 of the 15 highest-ranked predictors for dementia were common across the three models. Importantly, several novel predictors of all-cause dementia, such as health service utilization, were identified across these algorithms.
Additional authors include Kayleen Ports, a former UC Irvine master’s student, and Jiahui Dai, a current graduate student researcher, both from Wen Public Health; Kyle Conniff, a recent UC Irvine PhD graduate in statistics; and Maria M. Corrada, a professor of neurology in the UC Irvine School of Medicine. Spero M. Manson, a distinguished professor, and Joan O’Connell, an associate professor, with the Centers for American Indian & Alaska Native Health at the Colorado School of Public Health also contributed to the study.
