It was mostly her son who spoke. The woman sat in a plastic chair, hands clasped together in her lap, staring through us at something only she could see. Her son told us that she had been forgetting things for three or four years now. Her family took her to a doctor in Mumbai, who blamed her memory loss on age. No MRI. No diagnosis. No plan.
That morning in Karjat, my team and I captured retinal images of the woman and several others. That inexpensive image of her retina could enable AI-assisted screening for Alzheimer's disease well before a more expensive or invasive formal diagnostic test. I am developing such an Artificial Intelligence (AI) model and device. Research shows that because the retina and the brain originate from the same developmental tissue, the vascular and structural characteristics of the retina correlate with Alzheimer's-related processes in the brain. So, the retina is a window to the brain and the algorithms in trained AI models could recognise leading patterns of Alzheimer's from the retina well before the onset of cognitive degeneration symptoms.
Disease-modifying treatments for early Alzheimer’s, including lecanemab and donanemab, have received regulatory approval in the US, and donanemab launched in India in May. These treatments work best in early stages of the disease, highlighting the importance of early diagnosis. Two-thirds of dementia and Alzheimer's cases around the world are already in low- and middle-income countries, and in some of them as many as 90% of cases go undiagnosed. Current diagnosis methods need specialists combined with expensive PET scans, or invasive cerebrospinal fluid tests, or the increasingly practical p-tau217 blood test. Retinal imaging, if approved, could offer an alternative that is less biologically specific but potentially far more scalable, for screening rather than definitive diagnosis.
As I continued developing my AI model, a different central issue emerged. For AI models like mine to be accurate in the field with underserved populations, the model would need to be trained with data from these populations and not just the easily available Western datasets I had used. Machine learning algorithms learn patterns in the training data, and Indian and Western retinas differ in terms of pigmentation, vessel size, and branching patterns. A model trained on one population might be biased to classify normal features of another and overlook clinically relevant changes. Unlike other diversity problems, this is also an external validity problem and problem of the clinical method that limits how its findings can be reliably applied to a broader patient population.
{{/usCountry}}As I continued developing my AI model, a different central issue emerged. For AI models like mine to be accurate in the field with underserved populations, the model would need to be trained with data from these populations and not just the easily available Western datasets I had used. Machine learning algorithms learn patterns in the training data, and Indian and Western retinas differ in terms of pigmentation, vessel size, and branching patterns. A model trained on one population might be biased to classify normal features of another and overlook clinically relevant changes. Unlike other diversity problems, this is also an external validity problem and problem of the clinical method that limits how its findings can be reliably applied to a broader patient population.
{{/usCountry}}No company or institution should claim to provide population-level screening in India unless it has independently validated its model using Indian patient data in an Indian clinical setting. This holds true for other countries such as Nigeria, Indonesia, and Brazil. Regulators must mandate population-specific validation before approving any screening claims. The validation must include different demographic groups and in a realistic clinical setting, not just among specially selected research subjects. Evidence required to address this issue cannot be gathered after the development of screening methods on a non-representative population, and representation needs to be established as an explicit goal early in the Institutional Review Board (IRB) process. Requiring appropriate validation may add time upfront, but it can prevent greater delays later by identifying failures of generalisability before deployment.
Additionally, universities, funding bodies, hospitals and technology companies in wealthy countries have an ethical obligation to finance such work where possible, since much of it will be used or commercialised elsewhere. Where privacy, consent, and governance requirements permit, making appropriately de-identified datasets openly available could further maximise their scientific value and enable independent validation and new research.
Since I could not find any publicly available retinal dataset from India that included patients diagnosed with neuroimaging-confirmed Alzheimer’s disease, I am now working to construct the Indian Retinal Imaging Database for Alzheimer’s Disease (IRID-AD). This project involves the acquisition of 500 retinal images via hospitals and community health camps. The goal is to have 250 images from subjects with confirmed Alzheimer’s disease via neuroimaging techniques, and another 250 images from cognitively normal individuals. So far, we’ve acquired 100 images from 50 confirmed Alzheimer’s patients and 90 images from 45 control participants. Images have been taken after ethics approval, with proper consent, and de-identified prior to inclusion in the database. Camp participants were given a complete health evaluation, involving basic screenings such as blood pressure measurements and HbA1c levels. People who participate in scientific research deserve to gain a health benefit rather than being sources of data alone. The resulting database will be open access and freely available to any researcher.
The woman in Karjat may never know what happens to her retinal image. But one day, a researcher might analyse it together with hundreds of other images and obtain knowledge impossible to extract from any Western database.
Closing a data gap starts with collecting evidence from people whose health issues were left behind by the research agenda in the past. Sometimes, closing a data gap begins with a photo of one woman’s eye.
(The views expressed are personal)
This article is authored by Roshan Avik Srivastava, founder, ReCognEYE and principal investigator, IRID-AD, the Indian Retinal Imaging Database for Alzheimer’s disease.