When screening patients for dementia, incorporating standard retinal imaging into an assessment of traditional risk factors for the disease can improve detection of cognitive decline, investigators from the United Kingdom have reported.
The findings add to a growing body of evidence linking systemic disease and the eye —a field of research called “oculomics.”
“We show that deep-learning models trained with interpretable retinal images can be used in addition to traditional risk factors to act as a simple, noninvasive method for dementia screening,” said Robbert Struyven, MD, a doctoral candidate at the University College London Centre for Medical Image Computing, who presented the results at the 2023 annual meeting of the Association for Research in Vision and Ophthalmology. “We find that in addition to known predictors of dementia — such as older age, hypertension, and diabetes — signals from the retina can be used to help improve predictions.”
For dementia screening, the benefit of incorporating retinal information, in addition to traditional nonretinal risk factors, has not yet been fully established. The new study evaluated whether the addition of deep-learning algorithms trained on images generated by optimal coherence tomography (OCT) could boost the performance of dementia screening on top of the predictive power of traditional models of stratifying risk.
“The retina is an extension of the brain. That’s the impetus behind looking at OCT-based features in neurological disease,” said Ian Han, MD, an assistant professor of ophthalmology and visual sciences at the University of Iowa Carver College of Medicine in Iowa City.
The retina offers unique features that make it much more accessible than the brain for purposes of evaluation, Han added.
“In ophthalmology, we have detailed, high-resolution imaging right down to several microns in tissue,” he said. “We can see various features that cannot be seen very well on neuroimaging of the vasculature. So potentially, you could have a front end, detectable micro-scale feature that would be made more readily detectable than imaging, for example, with MRI.”
Four Models Evaluated
For the retrospective study, Struyven and colleagues examined data from 353,157 participants in AlzEye, a longitudinal record-level linkage of ophthalmic imaging and hospital admissions. The researchers identified patients diagnosed with Alzheimer’s disease, vascular dementia, or other forms of the cognitive decline, as well as cognitively healthy controls, and paired them with their ophthalmologic records.
Struyven’s group compared the performance of four screening models:
-
non-retinal risk factors
-
interpretable retinal features, such as vascular properties and the thickness of the retinal layers, along with traditional risk factors
-
deep-learning algorithms trained on OCT imaging
-
a multi-modal fusion model combining retinal images and traditional risk factors
Traditional risk factors included demographic factors — age, sex, ethnicity, and an index of socioeconomic status — and clinical features such as hypertension and diabetes linked to the development of dementia. Retinal features consisted of layer thicknesses on OCT and morphological features extracted from color fundus images. Separately, a computer algorithm was trained using retinal images to predict all-cause dementia. The features of the algorithm were used in the fusion model.
The models were trained on data from 36,877 patients visiting four Moorfields Eye Hospitals from 2008 to 2018 and were validated on two test datasets: 4083 patients visiting the same four hospitals and 4089 patients visiting three distinct hospitals.
“Quite a Jump”
As measured using an area under the receiver-operator curve (AUROC) analysis, the researchers found that the combination of deep learning, imaging, and risk factors was the most sensitive predictor of dementia (0.840 in the internal test and 0.805 in the external test).
“We saw a performance boost with the combination,” Struyven said. “It was quite a big jump.”
By type of dementia, the potential for detection was highest for Alzheimer’s disease (AUROC, 0.877), followed by vascular dementia (AUROC, 0.870) and other types of the condition (AUROC, 0.809).
However, as the underlying pathology of these diseases is different, separate models may be needed for each, the researchers noted.
“Our results provide further evidence for the possible benefit of using a noninvasive retinal exam in conjunction with traditional risk factors for population-based dementia screening,” Struyven said.
Struyven acknowledged the need for better external validation of their model and said his group plans to conduct additional studies to do so.
Daniel Ting, MBBS, PhD, an associate professor of ophthalmology at Duke-NUS Medical School in Singapore, observed that Asians comprised 23% of the training dataset but only 16% of the validation set. “Generalizability of models is important,” Ting said, pointing out that a model built on a Caucasian population may be unreliable when applied to Asians.
Potential Use in Screening
Ting also questioned the best setting for this potential multimodal screening model. “Will this technology be used in a community setting or a tertiary care setting, and will that be a neurology or ophthalmology setting? If a patient screens positive, what happens next? There are a lot of questions to answer,” he told Medscape Medical News.
The investigators said they envision their model, or one like it, to eventually be incorporated into some type of community-based screening for dementia.
Struyven, Han, and Ting report no relevant financial relationships.
Association for Research in Vision and Ophthalmology (ARVO) 2023 annual meeting: Abstract 1282. Presented April 24, 2023.
Caroline Helwick is a medical journalist with more than 25 years of experience reporting from medical conferences around the globe.
For more news, follow Medscape on Facebook, Twitter, Instagram, YouTube, and LinkedIn
Source: Read Full Article
