Artificial intelligence for diagnosing dementia
(Boston)—Advances in public health during the past few decades mean that more people worldwide are living into old age. As a result, there is a significant increase in the global burden of diseases commonly associated with aging, including dementia, especially Alzheimer’s disease. This combined with a projected physician shortage in coming decades, may limit the ability to provide timely care to those who need it.
A new study from Boston University School of Medicine (BUSM) researchers suggests that computational strategies (artificial intelligence/AI) can help to alleviate some of these difficulties of providing dementia care in an aging population. “Even in circumstances where a specialized neurologist or neuro-radiologist is busy to directly provide a diagnosis, it is foreseeable that some degree of automation could step in to help, thereby enabling doctors and their patients to plan treatment accordingly,” explains corresponding author Vijaya B. Kolachalama, PhD, FAHA, assistant professor of medicine at BUSM.
Past research has shown that AI models can make simplistic choices between “disease” or “no disease”, but that’s not how doctors treat patients. Instead, they must consider all possible conditions that could be affecting a patient in their clinic, relying upon physical examination, neuro-psychological testing, laboratory findings and imaging to uncover a unique signature that cinches the diagnosis. According to Kolachalama, this work comes much closer to this “real world” scenario, allowing a computer to focus on the true source of a patient’s illness even when there are different possibilities.
“We show that this is achievable when a model is presented with a broad differential diagnosis of possible illnesses. For context, “dementia” as we know it can be the result of different processes; the most common one being Alzheimer’s, but chronic alterations in a person’s mental status can also occur in other disorders – from Parkinson’s disease to geriatric depression to nutritional deficiencies and beyond. Our study is novel because, unlike work before it, we demonstrate a computational strategy for providing an accurate diagnosis during this diverse landscape of neurologic disease,” he adds.
The researchers designed a variety of computer models capable of digesting large quantities of data that might be collected during a typical work-up of a patient with suspected dementia, including results of neuro-psychological and functional testing, medical history, physical examination, demographics, and MRI scans. This information was then fed to a neural network which was then trained to elicit disease-specific signatures from this vast set of inputs.
Using specialized methods in machine learning, they were able to pinpoint the exact pieces of data that their model used in its diagnostic decision-making, including important neuro-psychological test scores, laboratory values and physical examination findings that could be suggestive of a specific disease . They then applied these same methods to localize dementia-related changes in MRI scans and found that the locations marked as “important” by the model corresponded to brain regions with microscopic evidence of degenerative tissue changes.
Lastly, an international group of physicians participated in a “head-to-head” comparative study with the AI models. Both the experts and the model were presented with an identical set of patients and asked to provide diagnoses using the same pieces of information. The accuracy of the doctors and the computer was similar.
Kolachalama believes that computational strategies can help to alleviate some of the difficulties of providing dementia care in an aging population. “In scenarios where patients may not be able to reach specialized neurologic care, our work could help to fill in the gaps and connect people with timely information about their health and the wellbeing of their loved ones.”
These findings appear online in the journal Nature Communications†
Funding for this study was provided by grants from the Karen Toffler Charitable Trust, the Michael J. Fox Foundation, the Lewy Body Dementia Association, the Alzheimer’s Drug Discovery Foundation, the American Heart Association (20SFRN35460031), and the National Institutes of Health (R01 -HL159620, R21-CA253498, RF1-AG062109, RF1-AG072654, U19-AG065156, P30-AG066515, R01-NS115114, K23-NS075097, U19-AG068753 and P30-AG013846).
Method of Research
Subject of Research
Human tissue samples
Multimodal deep learning for Alzheimer’s disease dementia assessment
Article Publication Date
Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.