AI in Ophthalmology: Three Applications in Risk Assessment, Detection, and the Fight Against Misinformation
Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. In this episode, we will talk about some of the areas of ophthalmology where artificial intelligence is being successfully applied.
Did you know that there is a very promising field of ocular biomarkers of systemic diseases called ''oculomics''?1
This new approach uses datasets from retinal imaging and artificial intelligence algorithms to bring incremental value to risk assessment scores in a range of systemic diseases.1
Given that retinal vasculature may be considered an indicator of systemic vascular health, impaired retinal vascular network patterns can be associated with increased cardiovascular risk scores, cardiovascular mortality, and cardiovascular disease. Here, oculomics could be a promising tool to add a personalized approach to traditional cardiovascular risk scoring.1
Moreover, it has been reported that traditional cardiovascular disease prediction models, such as the Framingham risk score for the calculation of 10-year risk of cardiovascular disease, as well as the recently updated SCORE2, may have some limitations in daily clinical practice for specific ethnic groups and patients with an intermediate risk profile.1
These limitations coupled with the continuous developments in imaging of the retina, such as retinal fundus photography, optical coherence tomography–angiography, or OCT-A, and even methods of adaptive optics, have prompted the search for automated analysis algorithms that can confirm previous findings about the association between characteristic retinal microvascular parameters and cardiovascular status.1
As reported in the literature review by Louis Arnould and colleagues in 2023, numerous algorithms have achieved a very high prediction rate and performance accuracy, as cardiovascular risk factors, patient stratification, and major cardiovascular events could currently be predicted with rates up to 80%.1
Let's review some of the most relevant findings from this review. First, the team led by Nusinovici developed in 2022 a deep learning algorithm called RetiAGE, which was trained with retinal fundus images to assess biological age based on retinal fundus. This is important because considering that aging is one of the main cardiovascular risk factors, the algorithm could be used in future studies for cardiovascular risk assessment.1
In addition, Kim and colleagues also found highly accurate age prediction with a convolutional neural network algorithm based on 24,366 fundus images. They reported that the differences between fundus-predicted age and chronological age were more significant after the age of 60 and in patients with systemic vascular diseases such as hypertension and diabetes.1
As early as 2018, Poplin and colleagues had designed a convolutional neural network model to predict multiple cardiovascular risk factors based on retinal fundus images from two datasets. The results were relevant, especially for predicting age, sex, current smoking status, and systolic blood pressure.1
Continuing this trend, Cheung and colleagues developed a deep learning algorithm for automated measurements of retinal-vessel caliber from retinal photographs. The model obtained comparable or better results than those of the experts in the association between central retinal arteriole and cardiovascular risk factors including blood pressure, body mass index, total cholesterol, and glycosylated hemoglobin levels.1
AI has also been studied in the prediction of biomarkers of cardiovascular disease by means of retinal imaging; this technology could enhance cardiovascular risk stratification based on historical risk score calculations.1
An interesting example is the estimation of coronary artery calcium, by means of deep learning models based on retinal photographs. Although calcium scores have shown great efficacy in patient stratification, its measurement remains invasive, expensive, and has a risk of radiation exposure for the patient. So, several studies have proposed estimating calcium scores using artificial intelligence algorithms based on retinal fundus images. Here are some of the highlights.1
First, we have the findings of Son's team, who presented a deep learning algorithm to discriminate between patients with high or low calcium scores based on retinal fundus photographs.1
While they showed a moderate area under curve of 0.823 with unilateral fundus images and 0.832 with bilateral fundus images; it is relevant that the discrimination was better with the usual cardiovascular risk factors.1
Interestingly, Son's team found that a better performance could be achieved with the combination of retinal fundus images and various traditional risk factors; namely age, presence of hypertension, and sex; which yielded an area under the curve of 0.886. Therefore, future algorithms based on retinal images should ideally integrate clinical characteristics to improve their performance.1
To aid in cardiovascular risk stratification and also in refining the classification of patient groups with borderline and intermediate risk scores on the Pooled Cohort Equations scale, Rim and colleagues developed in 2021 a deep learning-based calcium score prediction score, called RetiCAC.1
By studying fundus photographs, the RetiCAC's performance was comparable to that of cardiac computed tomography scans in predicting future cardiovascular events as an incremental feature to further stratify intermediate-risk groups, for example, those in the borderline risk group.1
On atherosclerosis, Chang's team focused on developing a prediction based on retinal fundus images, of abnormal carotid artery intima-media thickness as a surrogate marker for atherosclerosis measured by ultrasonography.1
Their model aimed to predict the deep-learning-funduscopic atherosclerosis score, known as DL-FAS, which was able to predict carotid artery atherosclerosis with an area under the curve of 0.713, allowing the researchers to refine the prediction made with the Framingham Risk score. Thus, the DL-FAS model could be added to the conventional risk stratification scores for a better longitudinal cardiovascular disease outcome prediction.1
Now moving on to actual retinal pathologies, where artificial intelligence algorithms have demonstrated exceptional accuracy in the classification of pathology from retinal fundus photographs in laboratory settings. But are these technologies ready to be used in ophthalmology clinics and other healthcare settings, such as family practices and emergency rooms?2
This question becomes more relevant now since the U.S. Food and Drug Administration has recently cleared two learning algorithms for screening for diabetic retinopathy and diabetic macular edema, respectively; making them the first machine-learning methods for diagnosis to be cleared by the FDA without the need of human oversight.2
A study by Pandey and team aimed to develop a stand-alone, independent retinal screening method capable of identifying multiple possible retinal pathologies. This method was constructed to be able to accomplish three basic tasks:2
1. To be capable of identifying multiple possible retinal pathologies.
2. To demonstrate performance equivalent to, or superior to the current standard of care in retinal diagnosis.
3. To produce reliable predictions that can be used by clinicians.
Seeking to meet these criteria, this team of researchers trained an ensemble-based deep learning algorithm capable of detecting three major retinal pathologies: diabetic retinopathy, glaucoma, and age-related macular degeneration, and to distinguish them from normal eyes.2
Now, for the methodology, first, 43,055 fundus images from 12 public datasets were used. These were split into 80% for the training set and 20% for the validation set. A separate test set was formed by randomly sampling the public datasets so that there were 25 images for each disease as well as for the healthy eye category, for a total of 100 test images.2
Furthermore, the ability of the model was compared with that of seven board-certified ophthalmologists with a practice duration range between 1 and 7 years, who used only information from the images to classify them, and were then asked to rank their confidence in their response.2
Compared to ophthalmologists, who achieved an average accuracy of 72.7% across all classes, the deep learning-based model achieved an average accuracy of 79.2%.2
Overall, the model showed both superior classification performance and reliability in comparison with the human experts, including a higher global mean specificity of 93.1% versus 90.9% achieved by the ophthalmologists. The model also had a higher mean positive predictive value, and a higher average global F1 score, which is a measure of the accuracy of the model.2
Interestingly, out of all of the images that were incorrectly classified by most ophthalmologists but correctly classified by the model, 54.5% of them were misclassified by ophthalmologists as 'normal' when they actually had mild diabetic retinopathy.2
According to the authors, the better performance of the model is a result of its ability to detect mild presentations of diabetic retinopathy in fundus images compared to ophthalmologists, since the data sets on which the model was trained contained a broad spectrum of diabetic retinopathy presentations.2
In addition, the test set included photographs of varying quality. Many of these would be considered suboptimal for the detection of retinal disease, and would be typically rated with less confidence by ophthalmologists.2
On the other hand, this model did not outperform ophthalmologists in groups such as glaucoma and age-related macular degeneration, where the number of training samples and original data sets was limited. The model's sensitivity in detecting age-related macular degeneration was 76%, which was considerably lower than that of ophthalmologists, who achieved an average score of 85.1%.2
In summary, this study demonstrated that it is possible to train a set of deep convolutional neural networks to accurately identify retinal pathologies and normal retinas from color fundus photographs. These advances could represent a breakthrough, since the automatic recognition of these diseases would increase efficiency in ophthalmology clinics, as well as introduce the potential of retinal screening in vulnerable regions where access to specialists is limited.2
Inicio de cápsula
Although clinicians often have access to additional patient data such as medical history, clinical examination, and auxiliary tests to make these diagnoses; these tests usually represent a high burden on human and technical resources.2
In current clinical practice, retinal images are manually interpreted by ophthalmologists. Automated AI classifiers may provide a method to rapidly screen populations for retinal diseases; thus, future works should explore the potential use of multi-disease artificial intelligence classifiers to leverage their efficiency, particularly in the social settings that need them the most.2
Fin de cápsula
Welcome back. In the previous segment, we examined the use of artificial intelligence on retinal images for both, traditional cardiovascular risk assessment, and detection of three important retinal pathologies, namely diabetic retinopathy, glaucoma, and age-related macular degeneration.
Now, we will discuss the impact of ChatGPT, an AI-based chatbot that has gained popularity as a source of information on a wide range of topics, including health and disease.3
ChatGPT was launched in December 2022, and was developed using supervised learning and reinforcement strategies to answer questions, including patients queries about disease or even treatments.3
To find out the accuracy of this artificial intelligence tool, a team of researchers led by Marie Louise Roed Rasmussen evaluated its responses to typical patient-related questions about vernal keratoconjunctivitis.3
As we know, vernal keratoconjunctivitis is a recurrent and severe allergic and inflammatory disease that primarily affects the pediatric population. Its management can be complex and costly, having an important psychological and socioeconomic impact. Its main symptoms include pain, burning sensation, itching, tearing, and sensitivity to light. It is common for patients or caregivers to have questions about the condition, but they may not always direct them to their physician, searching the internet for answers instead.3
Therefore, in this study, two experienced clinical experts in vernal keratoconjunctivitis asked 25 questions divided into four different categories:3
● General questions or etiology
● Treatment and/or prevention
Each question was asked consecutively five times, generating a total of 125 responses. Responses were rated on a Likert scale from 1 to 5 by two experts, with 1 being considered very poor or unacceptable inaccuracies; and 5 being very good, or without inaccuracies.3
As reported by the authors, the lowest scores were obtained in the areas of treatment and prevention, with questions such as:3
● How can vernal keratoconjunctivitis be treated?
● What types of surgeries are available for vernal keratoconjunctivitis?
● Can eye drops used in vernal keratoconjunctivitis have side effects?
● Are the eye drops used in vernal keratoconjunctivitis dangerous?
Some suggestions were inadequate, as calcineurin inhibitors were never mentioned, while some others were considered harmful by the experts, for example suggesting the removal of the conjunctiva.3
Even more, in the case of questions regarding the side effects of eye drops, the chatbot was unable to recognize the possible serious side effects of therapy; when asked about the safety of eye drops used in vernal keratoconjunctivitis, the answer did not include the possible serious side effects of topical corticosteroid use.3
The highest scores, on the other hand, were obtained in areas of general questions, etiology, and prognosis, such as:3
● Why do you get vernal keratoconjunctivitis mostly in spring and summer?
● What is the prognosis of vernal keratoconjunctivitis?
● Can you wear makeup if you have vernal keratoconjunctivitis?
Fortunately, many ChatGPT responses also included advice to contact a physician when experiencing side effects, or for other treatment or prevention queries. This tells us that, while it is important to be aware of their significant shortcomings, these types of chatbots could be tools that provide relevant answers, allowing us to better understand the information presented to the patients and their caregivers, and perhaps more importantly, to better understand the patients’ thoughts and decisions.3
Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode.
Discover more medical news on Viatris Connect.
- 1. Arnould L, Meriaudeau F, Guenancia C, Germanese C, Delcourt C, Kawasaki R, et al. Using Artificial Intelligence to Analyse the Retinal Vascular Network: The Future of Cardiovascular Risk Assessment Based on Oculomics? A Narrative Review. Ophthalmol Ther. 2023;12(2):657-674. Available at: Using Artificial Intelligence to Analyse the Retinal Vascular Network: The Future of Cardiovascular Risk Assessment Based on Oculomics? A Narrative Review
- 2. Pandey PU, Ballios BG, Christakis PG, Kaplan AJ, Mathew DJ, Ong Tone S, et al. An ensemble of deep convolutional neural networks is more accurate and reliable than board-certified ophthalmologists at detecting multiple diseases in retinal fundus photographs. Br J Ophthalmol. 2023; 2022-322183. Available at: https://bjo.bmj.com/content/early/2023/01/30/bjo-2022-322183
- 3. Rasmussen MLR, Larsen AC, Subhi Y, Potapenko I. Artificial intelligence-based ChatGPT chatbot responses for patient and parent questions on vernal keratoconjunctivitis. Graefes Arch Clin Exp Ophthalmol. 2023; doi: 10.1007/s00417-023-06078-1. Epub ahead of print. Available at: Artificial intelligence-based ChatGPT chatbot responses for patient and parent questions on vernal keratoconjunctivitis
Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.