Photo Credit: Jay_Zynism
At-home eye tests using AI and a smartphone accurately detect pediatric myopia, strabismus, and ptosis, enabling early referral for an eye evaluation.
An AI model developed to identify myopia, strabismus, and ptosis in children accurately identified these conditions using mobile photographs taken at home, according to findings published in JAMA Network Open.
“These results suggest that [the model] can assist families in screening children for myopia, strabismus, and ptosis, facilitating early identification and reducing the risk of visual function loss and severe problems due to delayed screening,” researchers wrote.
Harnessing AI for Timely Disease Detection
The study included 476 pediatric patients (>60% aged 6-12; >50% boys) who had been diagnosed with myopia, strabismus, or ptosis.
After generating 946 monocular photos to identify myopia and ptosis and 473 binocular photos to identify strabismus, researchers input the images into a deep learning network to independently detect myopia, strabismus, and ptosis.
The model performed well in identifying all three conditions, with similar results in boys and girls:
- Myopia: sensitivity, 0.84; specificity, 0.76; accuracy, 0.80
- Ptosis: sensitivity, 0.85; specificity, 0.95; accuracy
- Strabismus: sensitivity, 0.73; specificity, 0.85; accuracy, 0.80
Physician’s Weekly (PW) talked with Tamiesha Frempong, MD, MPH, and Mohamed Abou Shousha, MD, PhD, two ophthalmologists not involved in the study, about the potential impact of this technology.
PW: Why was it important to do this study?
Dr. Abou Shousha: Early diagnosis of pediatric eye conditions is essential, particularly because missing conditions like myopia, ptosis, or strabismus can lead to amblyopia, also called “lazy eye.” With amblyopia, the brain favors one eye due to untreated childhood vision issues, leading to permanent vision impairment if not addressed early. This study explores AI’s potential to provide scalable early screening, which is crucial for preventing amblyopia and ensuring timely treatment. However, this technology should not replace comprehensive clinical eye exams, which remain essential for diagnosing eye conditions with long-term consequences.
Dr. Frempong: Catching problems in kids early can change the trajectory of their visual acuity. It’s important to diagnose problems and initiate interventions early. Research and innovation beget more research and innovation. Studies like this one fertilize more innovative tools and resources that can lead to improved, timely patient care. These tools can help reduce the burden on healthcare systems with too many patients and too few providers. Studies like this also get creative juices flowing. They make you think, “How can I improve on this? What ideas can I contribute?”
What are the most important findings?
Dr. Frempong: As the authors note, while other research has explored AI and vision, their study is, to their knowledge, the first to use facial photographs to simultaneously predict myopia, strabismus, and ptosis in young patients. Although we’re just scratching the surface of AI, using this technology to screen pediatric patients can potentially decrease the burden on healthcare systems.
Dr. Abou Shousha: The study results are significant and encouraging. The high accuracy in detecting ptosis and myopia with basic mobile photographs demonstrates the remarkable potential of AI in enhancing clinical diagnostics. However, these are preliminary data, and further independent validation is required to fully assess the model’s performance in broader populations.
These results align with our progress at Bascom Palmer Eye Institute using AI and virtual reality technology to diagnose childhood keratoconus and strabismus.
Did the results surprise you?
Dr. Frempong: That the model was good at detecting myopia and ptosis and less good at detecting strabismus is not surprising because myopia and ptosis either are or are not present. But strabismus, unless it’s poorly controlled, may not always be present. Nor was it surprising when the model worked better with older kids because they could follow directions and cooperate better.
How might the findings affect patient care?
Dr. Abou Shousha: This study highlights the potential for AI tools to shift early detection into nonclinical settings, such as home screenings. With early detection, patients can receive timely referrals and interventions and have improved outcomes.
While AI may improve accessibility, clinical exams are still indispensable for confirming diagnoses and identifying conditions that require more detailed medical assessments.
Are any strengths or limitations of the study noteworthy?
Dr. Abou Shousha: The findings can potentially transform patient care by enabling earlier and more accessible detection of pediatric eye conditions using mobile phone technology, particularly in underserved regions. Integrating AI models into home-based screening could lead to quicker referrals and reduce the risk for amblyopia.
However, it is critical to emphasize that these tools should complement but not replace comprehensive clinical exams, which remain indispensable for fully diagnosing and managing eye conditions. These are preliminary data, and the lack of an independent clinical trial limits their generalizability. The results are encouraging, and the research team should plan for independent clinical trials to further advance their already impressive work.
Dr. Frempong: A significant limitation the authors acknowledge is that the model does not work as well with younger children, who most need to be diagnosed and treated to avoid irreversible vision loss. They acknowledge that the technology is binary, reporting a condition’s presence or absence but not its severity. The model doesn’t contextualize the results to say, “The condition is there, but does it matter?” So, using this tool in its current stage of development may lead to more, not fewer, visits to eye care providers. The technology needs to get smarter to tell us what is and is not clinically significant.
What questions remain unanswered for you?
Dr. Abou Shousha: Will this AI model perform equally well in larger and varied populations? I recommend further studies focusing on independent validation and examining the model’s ability to detect additional eye diseases.
Dr. Frempong: Participants needed to be able to follow instructions and cooperate. Studies like this can exclude the most vulnerable people: those who are neurodivergent, have ADHD, or are handicapped in some way or mentally challenged. Technology needs to improve to include the most vulnerable patients and to tell us whether a condition is present and important to pay attention to so we can triage our patients appropriately and treat the neediest first.
Is there anything else you’d like to mention?
Dr. Abou Shousha: While AI-based tools offer tremendous promise in early detection, they should be seen as complementary to and not substitutes for clinical eye exams. Comprehensive exams remain crucial, particularly for conditions that may have lasting effects on patients.
Dr. Frempong: Technology is amazing, but we need to ensure that we use technology as a tool and not allow it to supplant person-to-person doctor-patient encounters.
