With 108 million afflicted in 1980, an anticipated 425 million in 2017, and an expected 629 million in 2045, diabetes is on the rise and is now considered a global epidemic. In 2012, 93 million people were predicted to have diabetic retinopathy (DR), which is the main cause of blindness and visual loss globally. Of those, 28 million were estimated to have DR that was eyesight-threatening. Because early diagnosis and prompt treatment can stop the majority of vision loss from DR, developed nations have established DR screening programmes that focus on early diagnosis, monitoring, and prompt treatment of DR. These programmes frequently use telemedicine and rely on graders with specialized training analyzing fundus (back surface of the eye) pictures.

In the last ten years, artificial intelligence (AI) has advanced dramatically in the medical field, particularly in ophthalmology. By harvesting clinically pertinent information from medical data, machine learning (ML) and deep learning (DL) algorithms enable computers to manage or diagnose without direct human intervention. By easing the burden of visual impairment, enhancing DR results and patient satisfaction, lowering overall screening costs, and helping ophthalmologists lighten their workloads, AI systems used for DR screening must seek to improve population health overall. Additionally, AI systems should make it easier for people to get eye care, give diabetic patients the tools they need to control and improve their health, help people connect with their eye care team, and lessen the administrative and cognitive loads that both patients and the team face.

 Image- Medium

Modern machine learning techniques like Deep learning have demonstrated promising diagnostic results in the areas of speech recognition, image recognition, and natural language processing. It has shown solid results for medical imaging analysis in general in a number of medical specialties, including radiology and dermatology. Deep learning (DL), specifically for ophthalmology, carries on the long heritage of autonomous and assisted analysis of retinal pictures, which dates back to the 1990s and earlier. These artificial intelligence (AI) solutions have been shown to be more affordable, more accurate at making diagnoses, and more patient-accessible for DR screening. Recent studies on the use of DL in ophthalmology demonstrate how it could at least partially replace human graders while still maintaining a comparable degree of accuracy.

Challenges

Despite the immense potential of AI, there are unintentional, detrimental effects. Since AI has a pretty limited intelligence, it is designed to carry out extremely particular tasks on previously curated data from a single location. Since correlations are the foundation of most AI models, predictions made from health data may not generalize to other groups because of the variety of populations investigated. The difficulties with using AI for DR screening are listed below.

Guidelines for using AI in DR screening

Some policies to keep in mind

• AI designs should prioritize the demands of the end user and be human-centered.
• The ethical, social, economic, and legal consequences that may come from AI applications in DR screening should be addressed in clinical validation and transparency research, which should be given priority.
• AI used in DR screening should adhere to accepted safety, efficacy, and equity guidelines.
• The availability and affordability of AI systems for DR screening should be guaranteed by policy frameworks, and resources should be distributed fairly. To invest in infrastructure development, staff training, and development, validation, and maintenance of the AI system, payment and incentive policies must be in place.
• In order to keep frameworks adaptable to new opportunities and challenges, policy frameworks should stimulate stakeholder interaction and assist patient and consumer education regarding the usage and growth of AI in healthcare.
• Access to and usage of data should be made simpler by policy frameworks for the public, developers, and users of health AI technology.

Conclusion

Finally, expanding the capacity of eye screening within health systems has demonstrated enormous promise. This can be done by automatically classifying DR and other eye conditions across a variety of therapeutic applications. New obstacles to the integration of AI solutions in clinical care pathways and healthcare systems will inevitably arise as technology advances from clinical validation to translation with rising levels of technology readiness. For many of the identified problems, potential targeted remedies are already being researched and published. The ease of implementing AI may be increased by low-cost, simpler digital solutions, such as smart phone-based systems in resource-constrained environments. After taking everything into account, AI appears to be prepared for translation and implementation to change eye care.