With the help of artificial intelligence (AI) computer programs, researchers may one day be able to predict from medical records which people are more likely to develop Alzheimer’s disease later in life. According to an NIA-funded study, this may be possible by training certain self-educating programs — also known as machine learning algorithms — to spot risks via electronic health records. The results can then be used to identify the underlying genetic activity and biochemical pathways that elevate each person’s risk. These findings, published in Nature Aging, suggest that this new two-step approach may help both researchers and caretakers better understand the risks associated with each person’s case of Alzheimer’s.
Led by scientists at the University of California, San Francisco, the researchers first trained computers on the disease diagnoses, blood test results, and other important pieces of information stored in the electronic health records of 749 people with Alzheimer’s and 250,545 control individuals. Specifically, the researchers used random forest algorithms, which are types of computer programs that learn to recognize patterns in data.
Initial tests suggested that the programs are accurate at predicting Alzheimer’s risk strictly from electronic health records. The programs were about 70% accurate at predicting an Alzheimer’s diagnosis seven years before one happened and about 80% accurate at predicting one a year in advance. These predictions improved to around 86% to 90% accuracy when demographic — birth year, gender, race, and ethnicity — and visit-related information, such as first visit age and years in the health system, was incorporated into the programs. The results are comparable with similar studies.
In addition, the initial results suggested that certain pieces of information contained in the records may be strong predictors of an Alzheimer’s diagnosis.
Further training and testing showed that some factors, such as hyperlipidemia (high levels of blood fat), congestive heart failure, and arthritis, were consistently linked to an increase in Alzheimer’s risk, regardless of how far in advance the factors appeared in the records. Others, including dizziness, osteoporosis, and back pain, were important for predicting a diagnosis three years before, while mild cognitive impairment and vitamin D deficiency appeared to be important for predicting a diagnosis one year out.
The researchers also showed that the AI programs could be manipulated to identify male- and female-specific predictive factors. For example, osteoporosis appeared to be a warning sign for women while chest pain was one for men. Hyperlipidemia and congestive heart failure were predictive for both sexes, suggesting both conditions may play an important role in Alzheimer’s risk.
Finally, the researchers showed how certain genes and other biochemicals may play a role in linking the predictive factors with Alzheimer’s diagnoses. This involved the use of SPOKE, a special search engine that is designed to find previously published disease-related information stored in 21 different databases. One search showed that hyperlipidemia may be linked to Alzheimer’s via high levels of low-density lipoprotein cholesterol, a risk factor for heart disease measured in blood tests, and several genes, including APOE, a known genetic risk factor for Alzheimer’s. Another search found that some genes, including one called MS4A6A, may link some cases of osteoporosis to Alzheimer’s in women.
Overall, this study provides a blueprint for how researchers may use machine learning algorithms and other advanced computer programs to mine clinical and biological data to better understand each person’s risk for Alzheimer’s.
This research was funded in part by NIH grants R01AG060393, F30AG079504, P30AG062422, and T32GM007618.
NIA leads NIH’s systematic planning, development, and implementation of research milestonesto achieve the goal of effectively treating and preventing Alzheimer’s and related dementias. This research is related to Milestones:
- 1.H, “Enable better access to electronic health records data and provide support for their integration with clinical and molecular data to build person-specific predictive models of disease and wellness to enable disease sub-classification and inform data-driven drug repurposing.”
- 1.Q, “Support research that uses Artificial Intelligence (AI) and deep learning approaches for discovery and translational research.”
Reference: Tang AS, et al. Leveraging electronic health records and knowledge networks for Alzheimer’s disease prediction and sex-specific biological insights. Nature Aging. 2024;4(3):379-395. doi: 10.1038/s43587-024-00573-8.