9

In the year 2018, my father developed a strange complication that brought excruciating pain to his eyes. Even painkillers could not help. After several fruitless to numerous opticians, in desperation, I took to the internet for possible answers. I looked up the common visual illnesses that could lead to blindness and how to treat them but I couldn’t make a diagnosis since I was not a medical practitioner. Fortunately, a doctor discovered the infection and a few weeks later the pain was gone and my dad was healed. However, the thought of my dad almost going blind, and the changes it would bring to my family still haunted me years later. I could only imagine what families in my communities whose members suffer from blindness go through and when I acquired some skills in machine learning, I naturally decided to use them to try and solve a real-life problem around visual impairments. My research here led me to discover diabetic retinopathy.

 

In 2015, the International Diabetes Federation estimated that 2-5% of Kenya’s population has diabetes and over 50% of cases are undiagnosed. As the prevalence rises, cases of diabetic retinopathy also rise and greatly affect the population. In Kenya, 1 in every 3 people with diabetes develops diabetic retinopathy. Most patients seek treatment when it is too late and thus end up developing permanent organ damage to the eyes. Diabetic retinopathy is detected by frequent scans of the retina done by a certified medical practitioner. In Kenya, most people do not get regular medical checkups unless they feel sick. This causes a lot of cases that are reported to be detected late when anomalies are incorrigible and the damage too much to be reversed.

 

Different organizations and groups in the world have tried different solutions to try and reduce the number of cases of blindness caused by diabetic retinopathy. One of these solutions is employing artificial intelligence and deep learning in the detection of the disease. This involves training a computer model to analyze retina images from patients and point out those affected by the complication. Approximately 17% of the global population in 132 countries have access to less than 5% of the global ophthalmologist population. This shortage can be mitigated by setting up machines that require minimal supervision from ophthalmologists in more centers around world populations. The centers can diagnose and refer patients for treatment regularly before it is too late.

 

Bearing this in mind, I was motivated to develop, Smartsight, to provide a solution to this problem. I decided to leverage the power of the internet to help assist medical institutions and hospitals around the world. This project involves using artificial intelligence to aid the early detection of diabetic retinopathy in communities around me, thus reducing the number of cases of blindness. Smartsight is a software application that runs on cloud servers that are accessible over the internet. It utilizes Artificial Intelligence(AI), which refers to the simulation of human intelligence on machines that are programmed to think like humans, to make informed decisions when provided with a problem. The AI specifically employs neural networks where the application learns to perform certain tasks by analyzing training examples. In this case, the training examples will be sets of retina images with varying degrees of diabetic retinopathy progression from patients. Doctors will also specify the degrees of the progression of the disease to help train the AI model. Training, testing, and storing the models are computationally intensive therefore, these activities will be migrated to cloud resources where they can run faster and more securely. Privacy concerns arise when crowdsourcing confidential data from clinics in the country. Medical institutions and clinics in the country have strict policies regarding how patient information is shared or transferred. This is to maintain patient confidentiality and privacy. The Ministry of Health, however, sought to enable data sharing safely through guidelines in a project called mHealth. We will partner with clinics on the mHealth project to share data and all data being uploaded to the platform will be encrypted at the source before being stored. Encryption and decryption keys will be stored on the host’s computer. Grouped accompanying images will be available publically for use but patients have rights of ownership of their data and images and can redact them through the clinics. Access to the platform will be granted through a web interface portal. The trained model once deployed will be made available offline to participating clinics because most rural areas in Kenya do not have access to a fast internet connection. However, they will be required to update and sync the models with the cloud every 90 days in order to remain up to date.

 

According to the Community Eye Health Journal, there are about 50 practicing ophthalmologists in Kenya for a population of about 45 million people. This shortage of personnel along with other problems like deteriorating outreach services due to increased operating costs and shortage of equipment and supplies plus difficult procurement procedures make discovery and treatment of the disease difficult. This project aims to mitigate this by using artificial intelligence to help aid in the discovery of the disease. Unlike doctors, once an AI model has been trained it can be deployed to multiple locations at once and with little operating and maintenance costs. This should lead to an improvement in the discovery of affected patients early enough. Community members who choose to use these services will also be informed that they are not only helping themselves by sharing their data but also contributing to the wellness of others and the improvement of the Kenyan healthcare system. The project is going to be deployed to at least 20 optical imaging centers around Nairobi in the test phase. These centers are equipped with imaging machines that take pictures of the eye’s retina. Doctors can then upload the pictures to the artificial intelligence model and train it on what category the image falls. The application will ask the doctors to specify the degree of infection on a scale of 0-10. After a few months of training, we will proceed to the test phase where the accuracy of the model will be determined. If the accuracy is high enough we seek to expand to rural areas. The Kenyan government, through an initiative called Beyond Zero, set out to deploy mobile clinics in rural areas in 2013. These clinics availed essential maternal care services to patients in those areas. We plan on partnering with Beyond Zero in acquiring the scanning machines. These machines will be developed en masse and deployed in these clinics. The machines will use the deep learning model to determine if a patient has diabetic retinopathy. If the patient exhibits signs they will be referred to the closest specialist.

 

With this project, we aim to reach 300,000 people in the first year of its launch. From there we aim to increase that number by 100,000 every year until all mobile clinics are equipped and the majority of the population can receive check-ups regularly. Cases that will be discovered early can be treated and serious consequences such as blindness can be avoided. All patient data that is uploaded to the model is private and confidential, and as such should be treated with care on the internet. All user identities will be encrypted to prevent misuse of it and patients who do not want to share their information can retract it at any time from a clinic. This data will also be shared openly to help other organizations that might be interested in conducting research in this field do their research. The application has been fully open-sourced and we are thus accepting contributions from software companies and developers in the country to improve the application. With all these measures in place, we are going to provide an innovative solution to reduce the number of patients with diabetic retinopathy using a healthy internet.

REFERENCES

International Diabetes Federation. IDF Diabetes Atlas. 2015.

http://www.health.go.ke/site-data/uploads/2017/11/Guidelines-for-Screening-and-

Management-of-Diabetic-Retinopathy-in-Kenya.pdf

https://www.beyondzero.or.ke/download/beyond-zero-technical-report-2017-2018/

https://www.investopedia.com/terms/a/artificial-intelligence-ai.asp

http://news.mit.edu/2017/explained-neural-networks-deep-learning-0414

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1705981/

http://www.health.go.ke/site-data/uploads/2020/02/Revised-Guidelines-For-Mhealth-Systems-

May-Version.pdf

https://bjo.bmj.com/content/104/4/588