Artificial intelligence models hold incredible potential in the diagnosis, risk factor assessment, and prediction of the outcomes of myopia

The increasing prevalence of myopia is a global health concern, with high myopia increasing the risk of vision damage. This necessitates the use of artificial intelligence (AI) for early diagnosis, prevention, and management of myopia. Now, a Pediatric Investigation review sheds light on potential applications of AI in the early identification, risk assessment, and prevention of myopia. It further highlights the challenges and current development status of AI technology in the field of myopia.

Myopia, or nearsightedness, affects two billion people worldwide. Uncorrected myopia can impair vision, disrupting education, career prospects, and quality of life. By 2050, nearly half of the global population is estimated to become myopic. High myopia is often associated with complications that can lead to visual impairment, affecting patients’ quality of life and increasing the global medical and economic burden. Therefore, early diagnosis of myopia is important for the prevention of vision damage in patients with myopia.

Artificial intelligence (AI) has opened up new frontiers in the medical field and may be a solution to this global health concern. The subsets of AI, such as machine learning (ML) and deep learning (DL) can help analyze data to diagnose diseases and predict risk factors, biomarkers, and outcomes.

In a new literature review, Dr. Li Li, Dr. Jifeng Yu, and Dr. Nan Liu, all from the Department of Ophthalmology, Capital Medical University, China, have summarized the applications and challenges of AI in myopia, including detection, risk factor assessment, and prediction models. This study was published in the journal of Pediatric Investigation on 18 March 2025.

Interestingly, AI models can be trained using ML/DL to detect myopia from fundus photos and optical coherence tomography images. By feeding a model with a large quantity of fundus images from myopic patients, the AI can be taught to discern minute changes in color and pattern in the retina that are associated with myopia. This allows the model to diagnose future patients from their fundus photos.

In addition, self-monitoring equipment such as SVOne, a handheld device that uses a wavefront sensor to measure eye defects, can use AI algorithms to detect refractive errors in the eyes. The device could access an online database of images, which the AI can use as a reference to diagnose myopia. Moreover, AI can be trained to detect behavioral changes associated with the onset of myopia. Such detection is especially useful for the early detection of myopia in children, which is often ignored otherwise. For example, the Vivior monitor uses ML algorithms to note changes in visual behaviors, such as time spent on near vision activities, in children aged 6–16 years.

Furthermore, ML methods like support vector machine, logistic regression, and XGBoost can be employed to identify risk factors of myopia. “An XGBoost-based model can be fed large quantities of longitudinal data, allowing it to learn the outcomes and associated risk factors of myopia in numerous patients. This, in turn, allows the model to assess the risk factors of new patients based on their genetics, family history, environment, and physiological parameters,” explains Dr. Li Li.

Predicting the progression and outcome of myopia can help doctors adjust their clinical approach. Taken on a large scale, it can shape clinical practice and policymaking that help in myopia control. By feeding an AI model large quantities of biometric data, refractive data, treatment responses, and ocular images from numerous myopia patients, the AI can be taught to predict outcomes of myopia in new patients.

Despite the great potential of AI in myopia, several challenges need to be overcome. Firstly, it is important to ensure that the dataset used to train an AI model is correct and of high quality. Bias, false negatives/positives, and poor data quality can negatively impact the diagnostic and prediction accuracy of the model. Secondly, most AI models are trained using data from large hospitals, which may not be representative of patients going to smaller clinics. This creates a discrepancy between real-world and training populations. Thirdly, an AI model is not a trained doctor and may not be able to provide a clinical basis for its diagnosis, which can cause the diagnosis to be rejected by medical professionals. Finally, with such vast quantities of patient data used to train AI models, it is important to ensure the privacy of patients’ medical records.

“While our study highlights the remarkable progress made in the clinical application of AI in myopia, further studies are needed to overcome the technological challenges. By building high-quality datasets, improving the model’s capacity to process multimodal image data, and improving human-computer interaction capability, the AI models can be further improved for widespread clinical application,” concludes Dr. Jifeng Yu.