Doctors using the AI system correctly identified the responsible gene among their five most likely diagnoses in 88.5% of cases, compared with 67.3% among specialists who worked without AI support.

Polish scientists were among those involved in the research, including Professor Andrzej Grzybowski from the University of Warmia and Mazury in Olsztyn, head of the Institute of Ophthalmological Research in Poznań, and Adrian Smędowski, PhD, deputy head of the Department and Clinic of Ophthalmology at the Medical University of Silesia in Katowice.

Hereditary retinal diseases are particularly difficult to diagnose because more than 300 genes can be responsible for the conditions, while different mutations can produce similar clinical symptoms.

Diagnosis requires a combination of ophthalmological examination, multimodal imaging and genetic testing, which can be expensive and is not widely available.

Early diagnosis is increasingly important as gene therapies are developed for these conditions.

‘The sooner the most likely genetic basis of the disease is determined, the sooner the patient can be qualified for appropriate treatment or clinical trial’, Professor Andrzej Grzybowski said in a press release sent to the Polish Press Agency.

The AI system, called Retina4IRD, analyses fundus photographs, optical coherence tomography (OCT) images and basic clinical data to predict the most likely genetic cause of a disease before costly genetic testing is carried out.

The system was developed using data from 1,843 patients involving 3,376 eyes with genetically confirmed hereditary retinal diseases from nine centres in China, South Korea and Poland.

The researchers also conducted a prospective, randomised clinical trial involving nearly 300 patients with suspected hereditary retinal diseases. Physicians were randomly assigned to either use Retina4IRD or conduct standard diagnostics without AI support.

Both studies analysed 17 key genotype categories.

The study found that doctors using Retina4IRD achieved 88.5% accuracy in identifying the correct gene among the five most likely diagnoses, compared with 67.3% among specialists working without AI support.

Using the system also improved decisions about further diagnostics and therapeutic management.

According to Grzybowski, the AI system influenced doctors’ decisions. Physicians using Retina4IRD were more likely to change their initial diagnosis in a way that was consistent with the results of subsequent genetic testing, indicating that the system supported clinical decision-making.

Grzybowski said the system was not intended to replace genetic testing.

‘The system is intended to support diagnostics by identifying the most likely genotypes, while the final diagnosis still requires genetic confirmation. The system can therefore shorten the diagnostic process, reduce unnecessary genetic tests, help select appropriate genetic panels and accelerate the qualification of patients for gene therapy and clinical trials,’ he said.

However, Retina4IRD currently covers only the 17 most important genotypes. Most of the data also comes from Asian populations, while the system uses fundus photography and OCT without integrating other tests such as autofluorescence or electroretinography.

‘This study shows that artificial intelligence can support genotype prediction based on ophthalmic phenotype, and significantly improve the diagnostic process, which is a significant step toward precision medicine’, Grzybowski said.

He added that AI does not replace physicians but can streamline diagnostics and improve the accuracy of clinical decisions.

The research forms part of the activities of the Polish Consortium of Artificial Intelligence in Ophthalmology, established in 2024 at the initiative of Grzybowski.

The consortium brings together Polish clinical, academic and technological centres and international partners to collaborate on the development, validation and implementation of artificial intelligence methods for diagnosing and treating eye diseases. (PAP)

zbw/ agt/

tr. RL