Diagnosing an illness is easier than predicting how quickly it will get worse. That is the promise of artificial intelligence (AI) in healthcare: Helping doctors identify people who may be at a higher risk for a certain disease than others.

Researchers at Genetech and Roche have developed a deep learning model that may be able to predict which patients with vision problems related to diabetes are more likely to go blind.

The team just published a new paper in Nature Digital Medicine, “Deep learning algorithm predicts diabetic retinopathy progression in individual patients.” Their work is based on images of the inside of the eye.

With this new study, researchers wanted to develop a model that could predict which patients are at the highest risk of going blind within two years. Currently, doctors can estimate the progression risk for groups of patients with similar signs and symptoms, bu they can’t accurately predict the course of vision loss in an individual patient.

To train the model, the researchers used color photographs of the inside of the eye from patients with diabetic retinopathy. Researchers used deep convolutional neural networks (DCNNs) to assess the images and produce a target outcome prediction. DCNNs are often used for analyzing images. A DCNN assigns a degree of importance to various objects in the image and can differentiate one object from the other. With this retinopathy study, the algorithm was looking for hemorrhages and microaneurysms — symptoms of retinopathy.

This type of predictive algorithm could help patients receive individualized care. With this technology, ophthalmologists could identify high-risk patients and schedule more frequent monitoring visits.

“This deep learning technology paves the pathway for an AI tool that can inform management strategy with optimal check-up frequency and potential timely intervention to help preserve the vision of patients,” co-senior author Zdenka Haskova, M.D., Ph.D., a medical director in clinical ophthalmology at Genentech, said in a press release.

In addition to managing blood sugar levels, people with diabetes have to track other related health problems, such as kidney disease, foot and skin problems, heart disease, and vision problems. Almost all people with type 1 diabetes and about 60% of people with type 2 diabetes will develop retinopathy. In the early stages, a person’s vision gets blurry. The illness can progress without symptoms until a sudden loss of vision occurs. In the United States, about 7.7 million people have diabetic retinopathy. That number is expected to climb to 14.6 million in 2050.

The patient data in this study came from people already enrolled in two other clinical trials. The authors listed that factor as a limitation of the model because these patients didn’t necessarily represent the real world population of people with diabetes. Also, the researchers used a measurement scale that is mostly used in research settings, not physician offices. The other limitations include:

  • Small patient population at 530 people
  • Lack of an external validation set

The next step is to develop an algorithm that predicts vision loss directly.

Researchers also are using AI to improve automated insulin delivery systems. JDRF awarded $144,000 to a Swiss university to develop advanced algorithms that can predict dangerously low or high blood sugar levels. The goal is to optimize and personalize insulin treatment.

Another example of how AI and machine learning is changing standard treatment plans is Microsoft’s work with the Apollo Hospitals Group in Chennai, India. Dr. Sangita Reddy, the joint managing director at Apollo, described a project with Microsoft’s AI Network for Healthcare initiative. Launched in August 2018, the Cardiovascular Disease Risk Score API is designed to predict the risk of heart disease in the Indian population.

Researchers found that patients who had a normal check-up were having heart attacks. Risk assessments were not considering the right factors in estimating the heart attack risk. To improve the process of identifying patients at risk of an attack:

“The team started with 100 health check-up risk-factors and 200 lab data points and correlated each factor in relation to its significance to the occurrence of the disease. Eventually, they narrowed down to 21 risk factors to build model to predict heart risk for Indian population. Apollo Hospitals is looking at redefining how preventive health check-ups are done across its hospitals.”

Reddy also wrote that by combining applied AI and clinical expertise, physicians can help doctors make better treatment decisions for each patient.

Authors of the diabetic retinopathy paper are Filippo Arcadu, Fethallah Benmansour, Andreas Maunz, Jeff Willis, Zdenka Haskova and Marco Prunotto.

Researchers used these images of the inside of the eye to assess diabetic retinopathy severity. The image on the left shows a healthy eye and the image on the right shows signs of retinopathy.
Image: From the research paper, “Deep learning algorithm predicts diabetic retinopathy progression in individual patients,” published in Nature Digital Medicine, September 2019.