Model tested on tens of thousands of routine brain scans spotted disease risk with 90% accuracy
Although investigators have made strides in detecting signs of Alzheimer’s disease using high-quality brain imaging tests collected as part of research studies, a team at Massachusetts General Hospital recently developed an accurate method that relies on routinely collected clinical brain images. The advance could lead to more accurate diagnoses.
For the study, published in PLOS ONE, Matthew Leming, a research fellow at Mass. General’s Center for Systems Biology and an investigator at the Massachusetts Alzheimer’s Disease Research Center, and his colleagues used deep learning — a type of machine learning and artificial intelligence that uses large amounts of data and complex algorithms to train models.
In this case, the scientists developed a model for Alzheimer’s detection based on data from brain magnetic resonance images (MRIs) collected from patients with and without the disease who were seen at Mass. General before 2019.
Next, the group tested the model across five datasets — Mass. General post-2019, Brigham and Women’s Hospital pre- and post-2019, and outside systems pre- and post-2019 — to see if it could accurately identify Alzheimer’s based on real-world clinical data, regardless of hospital and time.
Overall, the research involved 11,103 images from 2,348 patients at risk for the disease and 26,892 images from 8,456 patients without Alzheimer’s. Across all five datasets, the model detected Alzheimer’s disease risk with 90.2 percent accuracy.
Among the main innovations of the work were its ability to detect Alzheimer’s regardless of other variables, such as age.
“Alzheimer’s disease typically occurs in older adults, and so deep learning models often have difficulty in detecting the rarer early onset cases,” Leming said. “We addressed this by making the deep learning model ‘blind’ to features of the brain that it finds to be overly associated with the patient’s listed age.”
Leming notes that another common challenge in disease detection, especially in real-world settings, is dealing with data that are very different from the training set. For instance, a deep learning model trained on MRIs from a scanner manufactured by General Electric may fail to recognize MRIs collected on a scanner manufactured by Siemens.
The model used an uncertainty metric to determine whether patient data were too different from what it had been trained on for it to be able to make a successful prediction.
“This is one of the only studies that used routinely collected brain MRIs to attempt to detect dementia. While a large number of deep learning studies for Alzheimer’s detection from brain MRIs have been conducted, this study made substantial steps toward actually performing this in real-world clinical settings as opposed to perfect laboratory settings,” said Leming. “Our results — with cross-site, cross-time, and cross-population generalizability — make a strong case for clinical use of this diagnostic technology.”
Additional co-authors include Sudeshna Das and Hyungsoon Im.
This work was supported by the National Institutes of Health and by the Technology Innovation Program funded by the Ministry of Trade, Industry and Energy, Republic of Korea, managed through a subcontract to MGH.
Source: The Harvard Gazette