New research published in Ophthalmology investigated the use of artificial intelligence (AI) models and smartphone-captured images to identify the early stages of specific pediatric eye diseases.
First, the basis for this research.
Investigators identified three ocular diseases that are commonly known to contribute to eye problems and potentially lead to long-lasting visual health damage among pediatric patients:
- Myopia
- Strabismus
- Ptosis
With the high prevalence of current and future global myopia cases already a known cause for concern, strabismus and ptosis can also have detrimental impacts on not just a child’s appearance, but also the development of their visual system.
Thus: Early screening and detection are critical for successful management and treatment of such ocular diseases—particularly among this young patient population.
Are there any obstacles with early screening and detection?
Most definitely, the investigators noted.
These obstacles are largely due to most forms of screenings being conducted in hospital settings by experts—leading to delays across the board, including for diagnosing and treating diseases.
Which leads to: The need for a quicker, more accessible form of screening that patients can conduct themselves.
Enter .. AI?
Exactly. AI and deep learning approaches have already been utilized for the early detection and potential diagnosis of a number of ophthalmic diseases—including inherited retinal diseases (IRDs) such as retinopathy of prematurity (RP) as well as diabetic retinopathy (DR), to name a few.
With this in mind: Researchers sought to develop an AI-based, multifunctional model that used mobile (ie: smartphone) photographs to predict the onset of myopia, strabismus, and ptosis in pediatric (child and adolescent) patients.
Let’s dive into this study.
Investigators conducted a cross-sectional study (Oct. 1, 2022 to Sept. 30, 2023) at the Ophthalmology Department of Shanghai Ninth People’s Hospital, School of Medicine, Shanghai Jiao Tong University in China.
The participants: 476 participants (aged 18 and under; 52.7% male) diagnosed with:
- Myopia
- Strabismus
- Ptosis
And the imaging component of this?
Participants’ faces were photographed under lighting conditions of 300 to 500 lux using a smartphone from a distance of 1.65in from each patient.
Then: Investigators used an AI-based deep learning network (ConvNeXt) to independently detect the three ocular diseases by inputting the photos into the network (cropping/resizing each image as needed, depending on the disease types).
Now the AI evaluation.
The researchers analyzed the AI model’s performance in identifying each of the diseases using the following assessment metrics:
- Sensitivity
- Accuracy
- Area under the curve (AUC)
- Positive predictive values (PPV)
- Negative predictive values (NPV)
- Positive likelihood ratio (P-LR)
- Negative likelihood ratio (N-LR)
- F1-score based on 5-fold cross-validations
Segmented age categories included 0 to 5 years; 6 to 12 years; and 13 to 18 years.
How else?
A GradCAM++ was also used to evaluate the weight of different regions within the eye images, as interpreted via a heatmap.
Note: Regions of greater significance were demonstrated by the use of warmer colors in the heatmap, with the key areas for identifying the diseases as follows:
- For myopia: The sclera (at the temporal edge of the pupil in the affected eyes)
- For strabismus: On the side of the affected eyes
- For ptosis: On the eyelids
So what was the patient breakdown of diseases?
Out of the 476 patients:
- Myopia (n = 251)
- Strabismus (n = 180)
- Ptosis (n = 171)
Note: A portion of patients had more than one or all three of the diseases (n = 7; 1.47%)
- Myopia + another disease (n = 107)
- Strabismus + another disease: (n = 99)
- Ptosis + another disease: (n = 44)
And the images captured?
A total of 1,419 photographs were evaluated and used to create AI-based, deep learning models for detecting all three ocular diseases. Of that number:
- 946 monocular photographs were utilized to develop a myopia and ptosis detection model
- 473 binocular photographs were utilized to develop a strabismus detection model
So … how accurate were these models?
Overall, the models demonstrated a high accuracy in detecting all three ocular diseases:
- Myopia: 0.80 (95% Confidence interval [CI], 0.78-0.81)
- Strabismus: 0.80 (95% CI, 0.79-0.82)
- Ptosis: 0.92 (95% CI, 0.91-0.93)
Performance for both models was noted as being constant across sex subgroups, with an increase in sensitivity for age subgroups (myopia and strabismus, specifically).
Talk specifics on these models.
First up: myopia.
- What the accuracy results mean: Essentially, 8 out of 10 patients could be correctly identified as myopic
- Even further: A greater prediction sensitivity was demonstrated in older age groups as well as more precise identification of myopia
- Translation: As the myopia level rose, the model’s ability to identify disease became easier
Next: strabismus.
- The model’s sensitivity: 0.73 (95% CI, 0.70-0.77)
- What this suggests: A demonstration of great capability to identify children and adolescents with strabismus accurately (including stable across the age subgroups), albeit lower than previous studies’ models.
Last: ptosis.
- The model’s sensitivity: 0.85 (95% CI, 0.82-0.87)
- What this suggests: Excellent screening capabilities, with notable variation in detection accuracy after subgroup analyses (age/sex)
Any notable limitations to consider?
Study authors noted a few:
- The study’s single-center, cross-sectional design with a small sample size
- As a result: There may be a need for a multicenter investigation to enhance the algorithm’s generalizability
- Use of only one captured photograph (rather than multiple from different perspectives) per patient
- As a result: This restricted the algorithm’s capabilities as a result of “insufficient information”
- Varying sample sizes for each disease (with myopia having the largest and strabismus the smallest) may have impacted each model’s detection sensitivity
See the complete list.
Lastly, what does this data indicate?
Per the investigators, AI prediction models utilizing smartphone photographs demonstrate a strong performance in accurately identifying ocular diseases in children and adolescents—providing a handy and early diagnostic tool for families to use at home.
Why this is key: With this capability, earlier identification can lead to a reduced risk of severe problems, including visual function loss, due to delayed screening.