New Machine Learning Model Detects Mild Cognitive Impairment Through Retinal Images
A recently developed machine learning model could analyze retinal images as well as associated data of specific features to recognize patients with mild cognitive impairment.
Sharon Fekrat, MD
Recently published in the journal of Ophthalmology Science, research on a developed machine learning model by researchers at Duke Health showed an ability to differentiate normal cognition from mild cognitive impairment (MCI) using retinal images. These findings highlight the potential for a noninvasive and inexpensive method to identify the early signs of cognitive impairment among patients that could possibly progress to Alzheimer disease (AD).1,2
The model analyzed 236 eyes of 129 controls and 154 eyes of 80 patients with MCI from retinal pictures and images as well as quantitative data to differentiate the 2 types of patients. The model’s ability to perform the MCI diagnosis reported sensitivity of 79% and specificity of 83% while achieving an area under the curve (AUC) of 0.809 when applied to an independent test set (95% CI, 0.681-0.937).
“This is particularly exciting work because we have previously been unable to differentiate MCI from normal cognition in previous models,” senior author Sharon Fekrat, MD, professor in Duke’s departments of Ophthalmology and Neurology, and associate professor in the Department of Surgery, said in a statement.1 “This work brings us one step closer to detecting cognitive impairment earlier before it progresses to Alzheimer dementia.”
Previously, Fekrat and colleagues created a model that used retinal scans and other data information from patients to successfully achieve a known diagnosis of AD. The scans, which are based on optical coherence tomography (OCT) and OCT angiography (OCTA), had detected structural changes in the neurosensory retina and its microvasculature in patients with AD.
READ MORE: Cognitive Function Sustained With Alzheimer Agent Bryostatin-1
“The retina is a window to the brain, and machine learning algorithms that leverage non-invasive and cost-effective retinal imaging to assess neurological health can be a potent tool to screen patients at scale,” colead author Alexander Richardson, a student in the Eye Multimodal Imaging in Neurodegenerative Disease lab at Duke, said in a statement.1
This recent study expands on the previous research with using machine learning techniques to identify MCI which is an often precursor for AD. The newer model detects specific features in the OCT and OCTA images that signal the presence of cognitive impairment, in addition to data from the patient. The patient data included age, sex, visual acuity, and years of education along with quantitative data from the images themselves.
In the analysis, 152 (64%) control eyes were used for training, 24 (10%) for validation, and 60 (25%) for testing. Subsequently, 104 (68%) MCI eyes were used for training, 20 for validation (13%), and 30 (19%) for testing. Authors reported no patient overlap between the training, validation, and test groups, and the patients were randomly assigned to each group. During the model evaluation on the test groups, individual eyes were randomly sampled so that the difference in ages between the MCI and control groups was not statistically significant from the previously outlined research.
“This is the first study to use retinal OCT and OCTA images to distinguish people with MCIfrom individuals with normal cognition,” colead author C. Ellis Wisely, MD, assistant professor in the Department of Ophthalmology, said in a statement.1 “Having a non-invasive and less expensive means to reliably identify these patients is increasingly important, particularly as new therapies for AD may become available.”
All told, the smaller sample size was a limitation of the study which contributed to the lack of an age-matched control cohort. Authors also excluded patients with known ocular and systemic diseases such as glaucoma, diabetes, and vitreoretinal pathology although noted that future work will include individuals with these diseases to expand the generalizability and more broadly differentiate those with MCI from those with normal cognition. Additionally, the size of the training set used limited the complexity of the model since using a bigger neural network with more or wider convolutional layers would risk overfitting, according to the authors.
The researchers noted that having a longitudinal dataset from individuals who have progressed from normal cognition to MCI to AD is needed for future studies. Overall, authors noted that this research provides further optimism around the primary goal with having the development of a model that can predict MCI or AD onset prior to the manifestation of clinical symptoms.
