Researchers from Pfizer and IBM claim to have developed a technique for machine learning that can predict Alzheimer’s disease years before symptoms develop. The team says their strategy achieved 71 percent specificity when evaluated against a group of cognitively stable people by evaluating small samples of language data collected from clinical verbal studies.
Alzheimer’s disease starts with ambiguous symptoms of mild memory loss, frequently misinterpreted, accompanied by a gradual, increasingly significant deterioration in cognitive abilities and quality of life. More than 5 million Americans of all ages have Alzheimer’s, according to the nonprofit Alzheimer’s Association, and every state is predicted to see at least a 14 percent increase in the prevalence of Alzheimer’s between 2017 and 2025. It is possible that the only way to postpone its development is by early intervention, because of the essence of Alzheimer’s disease and how it takes hold in the brain. But to keep it from accelerating, the disease is often detected too late.
IBM has previously investigated the use of AI to classify proteins that can predict amyloid-beta concentration, a peptide that shifts before memory disorders associated with Alzheimer’s are evident. And beyond IBM, some have investigated the potential of AI to detect Alzheimer’s and dementia hallmarks. For example, researchers at Unlearn. AI, a start-up that develops clinical testing software tools, recently published a paper setting out a framework that can predict the progression of the disease, predicting the symptoms that will be encountered by patients. Another paper co-authored by researchers at the University of California , Berkeley, describes an AI method that from brain scans up to six years before clinical diagnosis can ostensibly predict Alzheimer’s disease.
But IBM and Pfizer argue that this new study “significantly” differs from previous studies and the application of AI to help predict Alzheimer’s. The researchers worked with samples that were obtained before subjects in the study encountered the first signs of disability, in comparison to studies that predict initiation that rely on subjects displaying signs of cognitive disability. Instead of simply targeting high-risk populations, they have measured the risk of Alzheimer’s in the general population, capturing a variety of individuals, including those without a family history of the disorder or other risk factors.
The study contained 703 samples from 270 participants, half of whom had experienced signs of Alzheimer’s before 85 years of age. (The mean time to diagnose moderate Alzheimer’s was roughly seven-and-a-half years.) From a language perspective, over 87 variables including mispellings, punctuation use, uppercasing, verbosity, lexical richness, and repetitiveness were considered by the researchers. In addition, from the Montreal cognitive evaluation MoCA, they looked at age , gender, education, visuospatial and executive thinking, object naming, memory , attention, abstraction, and test results.
The IBM and Pfizer team analyzed the transcriptions of the samples of participants with natural language processing, allowing them to tap into AI to pick up subtleties and discourse shifts that they would otherwise have overlooked. And they drew on data from original participants (and their offspring and spouses) in the Framingham Heart Study, a population-based study overseen by the US, after receiving consent and approval from the Institutional Review Board of Boston University. Public Health Service to study cardiovascular disease epidemiology and threats. In the Framingham study, enrolled individuals are tested every four years with the two-minute Mini-Mental State Evaluation speech test and every year with neuropsychological evaluations when a family member experiences potential cognitive impairment.
These steps resulted in a larger dataset than those used in other studies and allowed real-life outcomes to validate predictions. For instance, if the model created by the co-authors of IBM and Pfizer predicted that a 65-year-old Framingham subject would develop Alzheimer’s by 85 years of age, they might review the records of that person to find out if the subject had been diagnosed with the disease and when the diagnosis took place.
Research has shown that much of the information used to train disease diagnosis algorithms can perpetuate inequalities. Recently, the U.K. team Scientists find that almost all datasets of eye diseases come from patients in North America , Europe, and China, suggesting that algorithms for the diagnosis of eye diseases perform less well for ethnic groups from underrepresented countries. Researchers at Stanford University reported in another report that much of the U.S. data for research concerning the medicinal use of AI comes from California , New York, and Massachusetts.
Indeed, within their own model , the researchers found evidence of bias, which predicted Alzheimer’s onset more accurately for participants without a college degree than for those with (76 percent versus 70 percent). In contrast with males, it also achieved greater accuracy with women (83 percent versus 64 percent), performing on average 2.61 times better than males for female subjects.
The IBM and Pfizer researchers, mindful of this, claim that as their work progresses, they expect to use datasets that build on the ethnic , social, and racial diversity of topics. In terms of disease prediction, this breadth of data is often very difficult to come by, and access to it has helped us to accurately train these models, “they wrote in a blog post.” “We [will] continue to train our algorithms while upholding the basic values of privacy, openness and consent at all times.”
The team claims that if their work were to ultimately hit production networks, which is published in The Lancet eClinicalMedicine, it could help doctors assess the need for more nuanced and demanding psychiatric evaluations, testing, and monitoring. This could also open the door to more successful clinical trials, as those considered highly likely to develop the disease could enter preventive therapy trials.
Our goal is to provide many AI and machine learning tools one day for clinicians to help determine whether an person is at risk of developing Alzheimer’s disease. [The accuracy of our model] is a substantial improvement over clinical scale predictions (59 percent), which is a forecast based on a patient’s other biomedical data available,’ the team continued. “One day, physicians will be able to use speech and blood tests together, using AI to assist them in assessing the likelihood of Alzheimer’s disease and setting the groundwork for preventive steps.”