AI-Powered Smartphone Diagnostics for Convergence Insufficiency

Abstract

Background: This study innovatively combines Artificial Intelligence (AI) algorithms with smartphone technology, automatically detecting the Near Point of Convergence (NPC) and diagnosing Convergence Insufficiency (CI) without the need for extra diagnostic tools and, notably, without having to rely on the subject’s vocal response, marking an unprecedented approach in the field to the best of our knowledge. Methods: This was a prospective study that enrolled 86 participants. The real-time tracking of eye structures and movements was conducted using AI technologies integrated with a mobile application (MobileS). Participants brought the smartphone closer, focusing on a target displayed on the screen. The system calculated pupillary distance (PD) and phone-to-face distance, incorporating a unique feature called the exodeviation episode’s counter (ExoCounter) to determine the NPC. Additionally, participants underwent testing using the RAF Ruler test (RulerT), considering the ground truth. Results: MobileS demonstrated significant correlation with the RulerT, as evidenced by a Pearson correlation coefficient of 0.74 (p < 0.001) and an Intraclass Correlation Coefficient (ICC) of 0.73 (p < 0.001), highlighting its reliability and consistency with conventional ophthalmic testing. Additionally, the system exhibited notable sensitivity and specificity in diagnosing CI. Notably, user feedback indicated a preference for the MobileS, with 71% of participants favouring it for its ease of use and comfort. Conclusions: MobileS is a precise, user-friendly tool for independent NPC measurement, applicable in tele-ophthalmology and home-based care. Its versatility extends beyond CI diagnosis, marking a significant advancement in ophthalmic diagnostics for accessible and efficient eye care.

1. Introduction

Accurately diagnosing ocular disorders is essential in ophthalmology. Advances in AI and mobile health are transforming diagnostic methods, addressing the rise in visual disorders linked to increased time and smartphone use [1,2,3]. New self-screening tools, including visual acuity tests, automated strabismus detection, and virtual reality (VR)-based assessments, are driving these advancements [4,5,6,7].

Tele-ophthalmology enables remote screening and consultation, improving diagnostic accuracy for conditions like diabetic retinopathy, age-related macular degeneration, and glaucoma while enhancing accessibility to care [8,9,10].

1.1. Convergence Insufficiency (CI)

Convergence is crucial for binocular vision and depth perception. Convergence Insufficiency (CI) affects near vision, causing eye strain, headaches, diplopia, difficulty reading, and blurred vision [11,12]. Its prevalence varies from 1.35% to 33%, depending on measurement methods [13,14]. Among fifth and sixth graders, prevalence reaches 13% [15], while in adults and military personnel with mild traumatic brain injuries, it ranges from 40% to 46% [16].

1.2. Measuring near Point of Convergence (NPC) and Convergence Insufficiency (CI) Diagnosis

Traditional NPC tests, such as the RAF Ruler test (RulerT) [17] and Pencil test (PencilT) [18], involve moving a target closer to the patient and recording when fixation is lost [19]. These methods are effective but require professional administration, rely on subjective patient feedback, and may be inaccessible in remote areas, leading to undiagnosed CI cases.

This study introduces MobileS, an AI-based system for automatic NPC detection and CI diagnosis using a smartphone’s front camera and the MediaPipe framework (Figure 1). MobileS eliminates the need for specialized equipment and professional supervision, making it accessible for patients with communication difficulties and enhancing tele-ophthalmology practices.

1.3. Purpose

This study compares NPC measurements from MobileS to those from RulerT and PencilT to assess its effectiveness in CI detection. RulerT remains the gold standard in CI diagnosis.

2. Materials and Methods

2.1. Measurements by Novel Algorithm

Our smartphone application integrates an AI algorithm to detect the NPC using the front camera. The algorithm extracts key features, including pupillary distance (PD), eye-phone distance, and eye movement relative to the bridge of the nose. These metrics are processed frame by frame, applying a mathematical technique known as Gaussian Smoothing to reduce noise while preserving data trends [21]. The final output includes the NPC, showcasing the algorithm’s efficacy in NPC detection and CI diagnosis. The MediaPipe system is designed to function across a wide range of smartphone models and camera qualities, ensuring consistent performance on different devices.

2.1.1. Step 1—Measurements of Static Eye Features and Distances

The algorithm uses the MediaPipe Eye Tracking Model, which employs an RGB camera to generate a 3D mesh of the face, focusing on the eye region [22,23] (Figure 1A). It detects 478 key points and processes them through a neural network to determine their 3D coordinates [24].

Key static parameters include pupillary distance (PD) (A in Figure 1C), white-to-white (WTW) measurements (B and C in Figure 1C), and distances from the eyes to the smartphone. The system tracks eye movements relative to each other and to the nose bridge (Figure 1C; the distance between B and D, and the distance between C and D).

The iris's diameter (WTW) is assumed constant at approximately 11.7 ± 0.5 mm [25,26,27], ensuring consistency in calculations.

The formula for PD (mm) is as follows:

$$\mathrm{P}\mathrm{D}={\displaystyle \frac{\left(\mathrm{R}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{E}\mathrm{y}\mathrm{e}\mathrm{C}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}.\mathrm{X}-\mathrm{L}\mathrm{e}\mathrm{f}\mathrm{t}\mathrm{E}\mathrm{y}\mathrm{e}\mathrm{C}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}.\mathrm{X}\right)\times 11.7}{\mathrm{I}\mathrm{r}\mathrm{i}\mathrm{s}\mathrm{D}\mathrm{i}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}}}$$

(1)

where PD denotes the pupillary distance, representing the distance between the centre of the eyes in millimetres. RightEyeCenter.X refers to the X component of the landmark representing the centre of the right iris in pixels, LeftEyeCenter.X represents the X component of the landmark denoting the centre of the left iris in pixels, and IrisDiameter is the diameter of the iris in pixels. The distance from the pupils to the bridge of the nose can be calculated using the same method as in Equation (1).

The smartphone-to-eye distance (Figure 1B) is calculated using

$$\mathrm{D}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{F}\mathrm{r}\mathrm{o}\mathrm{m}\mathrm{C}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{r}\mathrm{a} \left(\mathrm{d}\right)={\displaystyle \frac{\mathrm{F}\mathrm{o}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{L}\mathrm{e}\mathrm{n}\mathrm{g}\mathrm{t}\mathrm{h} \left(\mathrm{f}\right)\times 11.7 \mathrm{m}\mathrm{m}}{\mathrm{I}\mathrm{r}\mathrm{i}\mathrm{s}\mathrm{D}\mathrm{i}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}}}$$

(2)

where DistanceFromCamera (d) signifies the distance from the iris to the smartphone camera in millimeters, FocalLength (f) denotes the focal length of the smartphone camera in pixels, and IrisDiameter is the diameter of the iris in pixels.

As shown in Equation (2), this method, tested on various smartphone models, has shown an error margin of less than 10% [20].

2.1.2. Step 2—Measurement of Convergence

NPC is determined by analyzing changes in PD and smartphone distance. Normally, both decrease smoothly until fixation is lost (Figure 2A). If the PD stops decreasing or increases while the phone continues moving closer, the system detects a Break Point (Figure 2B). A threshold-based evaluation determines CI.

The ExoCounter, a novel feature, identifies simultaneous outward eye deviations linked to blurred vision [28]. By integrating ExoCounter with PD shifts (Figure 2C), the system enhances NPC detection and provides valuable insights into the patient’s convergence potential and mechanism.

2.2. Step 2—Clinical Trial

The clinical trial involves the validation of our MobileS compared to the established standard methods for measuring NPC and diagnosing CI and checking the reliability and accuracy of our algorithm.

2.2.1. Subjects and Facility

A prospective controlled study was conducted in accordance with the Helsinki Declaration and approved by the Ethics Committee.

Inclusion criteria: age 18 years or more who can sign the consent form, with best corrected visual acuity of 6/12 or more in each eye. Exclusion criteria: patients with a history of intraocular or strabismus surgery.

Participants used prescribed optical correction for accurate NPC measurements. Additionally, the MediaPipe framework used in our system supports eye tracking both with and without glasses, ensuring measurement accuracy.

2.2.2. Procedure

The NPC of each participant was assessed using three methods in a specific sequence: PencilT, RulerT conducted by a single researcher, and MobileS performed independently by the patient following a brief explanation. Between each test, participants were given a 3 min break, during which they were instructed to focus on a distant object. Each test was performed twice. Participants ranked usability post-assessment. The clinical trial involved validating the MobileS system by statistically comparing its NPC measurements to those obtained using the gold standard method (RulerT) for the same participants. This analysis assessed the reliability and accuracy of our algorithm. To ensure standardized execution, including movement speed, all NPC measurements in this study were conducted under professional supervision.

The RAF Ruler test (RulerT): An ophthalmologist moves a target closer until fixation is lost (due to blurred or double vision).

The Pencil test (PencilT): A manually held pencil replaces the ruler.

The Automatic System Test (MobileS): Participants follow a target on the smartphone until it touches their nose (Figure 3).

Historically and in conventional practice, an NPC of less than 10 cm (4 inches) is considered normal, as young people can maintain fusion at close distances [29,30]. If the NPC exceeds 10 cm, professional evaluation for CI should be conducted.

A population-based study found that the mean NPC was 8.59 ± 4.82 cm, with younger individuals (10–19 years) averaging 6.95 ± 3.87 cm. The NPC increased with age, reaching 13.06 ± 5.2 cm in those over 70 years [31].

In this study, we set 13.41 cm (8.59 + 4.82 cm [31]) as the threshold for CI while also considering the widely accepted 10 cm cutoff.

2.3. Statistical Analysis

The sample size (n = 86) justified the use of Pearson’s correlation coefficients to compare the measurements from our system and the traditional methods. Intraclass correlation coefficients (ICCs) were also calculated to evaluate the agreement between these methods. An ICC value close to 1 would indicate strong agreement.

3. Results

This study included 86 participants, consisting of 43 men and 43 women, with a mean age of 32 ± 11 years (ranging from 18 to 59 years). For calculating specificity and sensitivity, the RulerT was used as the gold standard.

Comparison of NPC Measurements: The mean NPC values measured by the PencilT, RulerT, and our MobileS were 7.57 ± 4.49 cm, 9.87 ± 3.04 cm, and 9.69 ± 3.59 cm, respectively.

Statistical Correlations: The statistical analysis revealed significant correlations between the methods. The Pearson correlation coefficient between RulerT and MobileS was 0.74, and between RulerT and PencilT was 0.81 (all p values < 0.001), suggesting a strong positive correlation in both comparisons (

Table 1. Correlations between the diagnostic methods.

Methods	N (Subjects)	r (Correlation)	p (Sig.)
RulerT vs. mobileS	86	0.74	<0.001
RulerT vs. PencilT	86	0.81	<0.001

Note: Pearson’s correlation coefficient test results.

). The intraclass correlation coefficient (ICC) between RulerT and MobileS was 0.73, and between RulerT and PencilT was 0.70 (all p values < 0.001), confirming a good level of agreement (

Table 2. Inter-device agreement by intraclass correlation coefficients (A, 1).

ICC (A, 1)	ICC Value	95% CI	p (Sig.)
RulerT vs. mobileS	0.73	(0.62, 0.81)	<0.001
RulerT vs. PencilT	0.70	(0.27, 0.86)	<0.001

Abbreviations: ICC (A, 1), intraclass correlation coefficients (two-way random effect model); 95% CI, confident interval.

).

Diagnostic Capability for CI: When the diagnosis of CI is based on an NPC threshold of >10 cm, the MobileS could identify CI in 26 out of 86 participants (30.23%), demonstrating a sensitivity of 76% and a specificity of 85% (Figure 4A). When CI was based on NPC > 13.41 cm, CI was found in nine patients (10.46%), corresponding to a sensitivity of 64% and a specificity of 94% (Figure 4B). A percentage of 83.72% of the NPC values obtained through the MobileS were within ±3 cm compared to the NPC values acquired by the RulerT (Figure 5).

Participant Preference: In terms of user experience, approximately 71% of participants favoured the MobileS, 21% preferred the RulerT, and 8% chose the PencilT. This indicates the acceptability and ease of use of our system in a clinical setting.

4. Discussion

Our study has successfully demonstrated the feasibility and reliability of the innovative MobileS for measuring the NPC and diagnosing CI. Additionally, it introduced a novel distinct feature (ExoCounter) alongside the conventional PD, allowing for the determination of NPC without supplementary instrumentation. Furthermore, the system facilitates self-testing by the patient. MobileS demonstrates high correlation and agreement with the RulerT and has received positive feedback from users for its ease of use. Notably, the questionnaire completed by participants at the end of the test revealed that approximately 71% of respondents preferred the MobileS system over the other two methods. We expect this percentage to increase as the prototype evolves with further development and as users gain more experience with home-based testing.

4.1. Unique Aspects and Clinical Significance of MobileS

MobileS uses AI and a smartphone camera for remote NPC assessment, eliminating the need for specialized equipment or in-clinic supervision. This enhances accessibility and allows for broader population screening. The system works with any cellular or Wi-Fi network, making it effective even in remote areas where VR-based CI tools may be impractical. Additionally, the growing integration of VR and LTE-enabled devices opens future opportunities for compatibility with CI treatment platforms [32,33]. MobileS also provides immediate diagnostic feedback, enabling users to track progress and share results with healthcare providers—an important advantage in tele-ophthalmology and rehabilitation.

This study has some limitations. In unsupervised settings, the application currently lacks real-time monitoring to correct potential testing errors, such as inappropriate testing speed, gaze deviation, or poor lighting conditions. Additionally, children, a key clinical target group, were not included in this study. Lastly, while the sample size is reasonable, further expansion would enhance generalizability and statistical robustness. Future improvements may involve algorithmic modifications to address occasional inaccuracies, such as test performance or gaze deviations.

Regarding sensitivity, it is important to note that our system is compared to the RulerT, which is the gold standard with 100% sensitivity. Therefore, our objective is to achieve sensitivity as close as possible to that of the RulerT. When compared to the PencilT, MobileS exhibits higher sensitivity. Specifically, at a threshold of 13.41 cm, MobileS achieves 64% sensitivity compared to 36% for the Pencil Test. At a threshold of 10 cm, both methods show 76% sensitivity. Thus, MobileS demonstrates superior sensitivity as a diagnostic tool compared to the Pencil Test.

We initially established normative data thresholds based on literature but have collected demographic data (such as age and gender) to refine these thresholds if needed. Future improvements may involve adjusting algorithms to address occasional inaccuracies, such as variations in test performance or gaze deviations.

While this study focuses on diagnosing CI using MobileS, it does not include post-treatment data. However, a follow-up study is underway to compare app-based convergence exercises with traditional pencil push-ups, aiming to assess clinical benefits. Although MobileS provides an accessible screening tool, clinical referral remains essential for confirmation and treatment management. The system is designed to recommend follow-up evaluations for those requiring further assessment.

4.2. Future Work

We continue to refine NPC detection by identifying true convergence points and distinguishing them from “NPC Mimickers” based on previous research [34]. Incorporating machine learning, we aim to improve accuracy by analyzing eye behaviour near NPC. While this study does not yet utilize training and test sets, future versions will integrate machine learning to enhance precision.

Planned expansions include detecting binocular vision anomalies such as strabismus and nystagmus, assessing accommodation, and providing vision therapy exercises. These enhancements align with our current technology, which supports both diagnostics and rehabilitation. Key challenges involve adapting the algorithm for diverse eye movement patterns and collecting high-quality datasets for validation.

Currently, MobileS is available for Android, with plans to extend support to iOS.

5. Conclusions

MobileS stands as a transformative tool in ophthalmic diagnostics, heralding the era of telemedicine and home-based care. Its AI-driven approach and smartphone integration provide accessible, efficient, and innovative eye health solutions. Given its accuracy and user-friendly design, MobileS has the potential to expand its role in diagnosing a range of ocular disorders, marking a shift in modern ophthalmology.

6. Patents

This work is protected by U.S. Patent Application CRML-P-047-USP, titled “System and Method for Diagnosis and Treatment of Various Movement Disorders and Diseases of the Eye”. Please note that the patent application is currently pending and has not yet been granted.