Machine Learning-Based Prediction of Strabismus with and without Amblyopia from Fixation Eye Movements

Abstract

Objective

To evaluate whether machine learning (ML) models trained on fixation eye movement (FEM) features can differentiate strabismic amblyopia, strabismus without amblyopia, and visually normal controls.

Design

Cross-sectional study.

Participants

A total of 159 subjects were recruited from a single Cleveland Clinic site: 49 with moderate/severe strabismic amblyopia, 24 with treated/mild strabismic amblyopia, 47 with strabismus without amblyopia, and 39 controls.

Methods

Fixation eye movements—horizontal and vertical eye-position traces from both eyes—were recorded during monocular viewing: fellow-eye and amblyopic-eye viewing for amblyopic subjects and right-eye and left-eye viewing for strabismic subjects without amblyopia and controls. Fixation and vergence instability, fast and slow FEM amplitude and velocity, fast-FEM frequency, and time-based eye deviation control metrics were computed using density-based clustering scans. Features were categorized as monocular, interocular within-viewing (nonviewing vs. viewing eye), interocular across-viewing (comparing the viewing eye across fellow-eye viewing vs. amblyopic-eye viewing or right-eye viewing vs. left-eye viewing), and binocular-coordination metrics (vergence instability and eye deviation measures). Six binary classification tasks were evaluated and trained using multiple ML models. Shapley Additive exPlanations (SHAP) identified the most predictive features.

Main Outcome Measures

Accuracy, area under the receiver operating characteristic curve (AUROC), and SHAP-derived feature importance for each classification task.

Results

Classification was strongest in distinguishing moderate/severe disease: amblyopia versus strabismus: accuracy 0.81 ± 0.09, AUROC 0.85 ± 0.09; strabismus versus controls: accuracy 0.90 ± 0.07, AUROC 0.92 ± 0.07; amblyopia versus controls: accuracy 0.90 ± 0.07, AUROC 0.97 ± 0.03. Performance was reduced for treated/mild amblyopia versus strabismus (accuracy 0.70 ± 0.12, AUROC 0.60 ± 0.19) but remained strong for treated/mild amblyopia versus controls (accuracy 0.94 ± 0.06, AUROC 0.87 ± 0.13). Classification of all amblyopia and strabismus combined versus controls achieved an accuracy of 0.89 ± 0.05 and AUROC 0.93 ± 0.04. Shapley Additive exPlanations analysis showed that abnormalities in the viewing amblyopic eye relative to the viewing fellow eye were the strongest predictors of amblyopia, whereas nonviewing-eye FEM metrics and binocular-coordination features best predicted strabismus with or without amblyopia compared with controls.

Conclusions

This proof-of-concept study demonstrates that FEM-derived features combined with ML can objectively distinguish strabismic amblyopia from strabismus and controls, with strongest performance in moderate/severe disease. Future work will extend validation to younger populations and incorporate portable eye-tracking devices to support scalable, point-of-care screening.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Artificial intelligence (AI) is increasingly shaping ophthalmology, with automated systems achieving high accuracy in screening for diabetic retinopathy, glaucoma, and macular degeneration.1, 2, 3, 4 However, progress in pediatric ophthalmology has been comparatively limited despite its urgent need.5, 6, 7 Disorders in early childhood emerge during a critical stage of visual development, yet their diagnosis often depends on subjective assessments that can be challenging to perform in pediatric populations.8, 9, 10 These challenges make pediatric eye care an ideal candidate for objective, technology-driven solutions. Current AI applications in pediatric ophthalmology have primarily focused on retinopathy of prematurity and congenital cataracts, with emerging work in refractive error.^5,11 Yet, these efforts remain sparse and have not addressed critical diagnostic nuances of common childhood eye diseases that influence treatment outcomes. One such distinction involves strabismus and amblyopia, 2 common pediatric conditions that often co-occur. Strabismus, characterized by ocular misalignment, disrupts normal binocular vision and may lead to amblyopia—reduced visual acuity in one eye during the critical period of visual development.^6,10,11 Untreated amblyopia results in permanent visual deficits despite surgical or optical correction of alignment, with long-term consequences for education, employment, and psychosocial well-being.^12,13 Early detection and treatment are therefore essential; however, clinical diagnosis remains challenging because conventional methods rely on visual acuity measurements, which are often unreliable in young or preverbal children.¹⁴ Alternative approaches, such as preferential looking tests, are examiner-dependent and lack robust test–retest reliability, contributing to underdiagnosis and delayed intervention.^15,16

Fixation eye movements (FEMs)—small involuntary movements during attempted fixation—offer an objective biomarker of visual function. Physiologic FEMs, comprising fixational saccades and intersaccadic drifts, prevent neural adaptation and visual fading while maintaining foveal image stability allowing achievement of best possible visual acuity.17, 18, 19, 20, 21, 22, 23, 24, 25 Prior work has demonstrated systematic increases in fixation instability (FI) and fixational saccade amplitude in amblyopia, with severity correlating with visual acuity deficit of the amblyopic eye (AE).^18,26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36 Strabismic subjects also exhibit increased instability, but without the interocular asymmetry observed in amblyopia, where greater instability occurs under amblyopic-eye viewing (AEV) compared with fellow-eye viewing (FEV).37, 38, 39, 40 Strabismus with amblyopia exhibits increased variability in both ocular deviation and vergence instability across different monocular viewing conditions compared with strabismus without amblyopia.³⁶ These differences provide a clear basis for distinguishing combined strabismus-amblyopia cases from isolated strabismus without amblyopia.

Recent advances in portable eye-tracking technology and machine learning (ML) have made pediatric eye movement assessment increasingly feasible.^5,7,41, 42, 43, 44 Convolutional neural networks applied to gaze deviation images have achieved high accuracy in strabismus detection.42, 43, 44, 45 Algorithms have also been developed for automated screening of amblyogenic risk factors using Bruckner pupil red reflex imaging.⁴⁶ In contrast, the use of FEMs for detecting amblyopic and strabismic pathologies remains limited.⁴¹ Our group previously developed an attention-based deep learning model capable of distinguishing anisometropic and strabismic amblyopia from control subjects using raw, calibrated eye-position traces obtained after user-guided preprocessing. Notably, the model did not rely on isolated FEM-specific features; instead, it was trained directly on the unparsed eye-position data.⁴¹ Despite these advances, prior AI research has not addressed the clinically critical distinction between strabismus with amblyopia and strabismus without amblyopia. This study aims to establish a proof-of-concept for leveraging FEM-derived features within an ML framework to enable this distinction, with the long-term goal of developing accurate, point-of-care screening tools that facilitate timely intervention during the sensitive developmental window when visual outcomes can be optimized.

Methods

Study Design and Ethics

This study is a retrospective observational study using data collected from a protocol approved by the Cleveland Clinic Institutional Review Board. Written informed consent was obtained from each participant or parent/legal guardian, as mandated by the Declaration of Helsinki. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD + AI) reporting guidelines were used in the development of this study.⁴⁷

Study Data and Participants

This study recruited 24 subjects with co-existing treated/mild amblyopia and strabismus, 49 subjects with moderate/severe amblyopia and strabismus, 47 subjects with strabismus without amblyopia, and 39 healthy controls at a single Cleveland Clinic site. All participants underwent comprehensive eye examinations, including visual acuity and stereoacuity measurements using the Titmus Stereoacuity Test, cycloplegic refraction, and strabismus and ocular motility assessment. Amblyopia was defined as reduced corrected distance visual acuity without structural abnormalities of the optic nerve, retina, or visual pathways. Subjects in the amblyopic cohort demonstrated (1) at least a 2-line difference in visual acuity between the 2 eyes at the time of diagnosis and had a history of strabismus defined per previous Pediatric Eye Disease Investigator Group studies as heterotropia at distance (with or without spectacles) and/or (2) documented history of strabismus that is no longer present.⁴⁸ Subjects with amblyopia were classified as treated if the visual acuity in the worse eye is ≤0.09 logarithm of the minimum angle of resolution (logMAR), mild if the visual acuity was between 0.09 and <0.30 logMAR, moderate if the visual acuity was between ≥0.30 and <0.70 logMAR, and severe if the visual acuity was ≥0.70 logMAR per previous Pediatric Eye Disease Investigator Group studies.⁴⁹ Strabismus without amblyopia subjects had neither current amblyopia nor any history of treated amblyopia.

Control subjects were selected based on the absence of ocular or systemic conditions affecting visual function, apart from refractive error, and were allowed to have small, asymptomatic phorias within established normative limits (∼1 prism diopter exophoria at distance and ∼ 2–4 prism diopter exophoria at near);⁵⁰ individuals with manifest strabismus, decompensated phorias, or symptomatic binocular dysfunction were excluded. The exclusion criteria included any coexisting ocular or systemic disease, congenital infections or malformations, and developmental delay. We also excluded subjects with idiopathic infantile nystagmus, characterized by an increasing-velocity slow-phase, or alternatively, pendular nystagmus.^51,52

Study Outcomes

The primary classification task was the discrimination of moderate/severe strabismic amblyopia from strabismus without amblyopia. Secondary classification tasks included discrimination of moderate/severe strabismic amblyopia from controls, strabismus without amblyopia from controls, treated/mild strabismic amblyopia from strabismus without amblyopia, treated/mild strabismic amblyopia from controls, and all amblyopia and strabismus combined from controls.

Eye Movement Recordings, Data Analysis, and FEM Features (Predictors)

Eye movement data were acquired with the EyeLink 1000 Plus system (SR Research), a high-accuracy, video-based eye tracker, following previously established protocols.^27,53 Participants used their habitual optical correction, when needed, to ensure optimal visual acuity throughout testing. Each participant was seated comfortably in a dimly lit room, with head position stabilized via a chin rest placed 84 cm from a liquid crystal display presenting the visual targets. Monocular calibration and validation were carried out using a 5-point array of horizontally and vertically arranged fixation targets, adhering to the manufacturer’s standard procedure. Following successful calibration, FEMs were recorded as participants maintained gaze on a white circular stimulus (0.5° visual angle) presented against a black background on a 30-inch liquid crystal display monitor (2560 × 1600 resolution, 60-Hz refresh rate, brightness 350 cd/m²). Each participant completed two 45-second monocular trials presented in random order: (1) left-eye viewing (LEV) for control participants and strabismic subjects without amblyopia or FEV for strabismic participants with amblyopia. (2) Right-eye viewing (REV) for control participants and strabismic subjects without amblyopia or AEV for strabismic individuals with amblyopia. An infrared-transparent filter occluded visible light to the nonviewing eye while still allowing the system to track it. Throughout testing, gaze position was monitored continuously to ensure stable fixation, and eye-tracking software recorded eye-position data for the duration of each trial. To help maintain engagement in pediatric participants, an animal sound was automatically triggered whenever gaze drifted outside a 4° × 4° fixation window for >500 milliseconds. Trials were monitored in real time for data quality. Periods with gaze deviations exceeding the 4° fixation window were flagged, and trials were repeated if <30 seconds of usable fixation data were obtained. This occurred in ∼10% of subjects. This approach ensured sufficient analyzable data while maintaining feasibility, particularly in younger participants.

Eye Movement Data Analysis and FEM Features

Eye-position data were processed in MATLAB (MathWorks). Blinks and partial blinks were detected and removed using previously established criteria.²⁷ Eye velocity was obtained by differentiating the position signal and applying a Savitzky–Golay filter to reduce high-frequency noise. Fixation stability was assessed using the bivariate contour ellipse area (BCEA) which encompasses 68% of fixation points.^54,55 Bivariate contour ellipse area was computed analytically from the covariance matrix of the fixation data, assuming a bivariate normal distribution. The BCEA is calculated using the formula: $\text{BCEA}=2k\pi {\sigma }_x{\sigma }_y{\left(1–{\rho }^2\right)}^{0.5}$

where k = 2.291 is the χ² value (2 degrees of freedom) corresponding to a probability of 0.68; σ_x and σ_y are the standard deviations of horizontal (x) and vertical (y) eye position of the AE, respectively; and ρ is the product–moment correlation of 2-position components (i.e., the horizontal [x] and vertical [y] position components). Bivariate contour ellipse area values were computed for both the right and left eye under monocular viewing conditions. Vergence instability was measured by calculating the difference in horizontal and vertical positions between the 2 eyes at each time point, using the same BCEA formula.^27,56 A log10 transformation was applied to normalize the BCEA values, with higher values indicating greater instability.

Fixation, interocular, and binocular-coordination metrics (vergence instability and eye deviation measures) were then analyzed for all study participants. To illustrate the interpretation of these metrics, representative examples are shown in Figure 1, which illustrates fixation of left/fellow eye (blue) and right/amblyopic eye (red), and vergence instability (purple) of 3 subjects: (1) control, (2) strabismus without amblyopia, and (3) strabismus with amblyopia across FEV/LEV and AEV/REV conditions. Blue and red values correspond to BCEA-based FI for the left/fellow and right/amblyopic eyes, respectively; purple values represent vergence instability. Interocular FI ratios (black numbers) were derived by dividing the BCEA of the amblyopic or right eye during AEV/REV by that of the fellow or left eye during FEV/LEV. Smaller ratios reflected more stable fixation. The subject with amblyopia had greater FI in the AE compared to right eye of strabismic subject without amblyopia and control subject. In addition, there was a greater interocular difference in FI across viewing conditions in amblyopic subject where the FI in the AE in AEV (2.40) was worse than FI in fellow eye (FE) during FEV (1.26) (panel c, blue boxes), compared with the interocular fixation difference between right eye during REV (1.22) and the left eye during LEV (1.17) in the subject without amblyopia (panel b, red boxes) and control (right eye during REV [0.63]; left eye during LEV [0.71]) (panel a, green boxes). The interocular FI ratios across viewing condition were increased in amblyopic subject compared to the other 2 subjects.

Figure 1.

Fixation instability (FI), interocular FI ratio, and vergence instability (VI) from a (A) control, (B) strabismic subject without amblyopia, (C) strabismus subject with amblyopia under fellow-eye/left-eye viewing (FEV/LEV; left column) and amblyopic-eye/right-eye viewing (AEV/REV; right column). Fixation and vergence stability are quantified using 68% bivariate contour ellipse area (BCEA) (red ellipse), with higher log BCEA values and larger ellipses showing higher log BCEA values indicating greater instability. The red and blue numbers represent the BCEA values indicating fixation instability for each eye, and the purple numbers represent the BCEA values indicating vergence instability. The black numbers represent the interocular fixation instability ratio across viewing conditions, computed based on fixation instability of the AE or RE in AEV or REV divided by FE or LE in FEV or LEV condition. The lower numbers equate to more stability; higher numbers equate to more instability. The blue boxes represent FI in the viewing eye, that is, AE in AEV and FE and FEV in subject with amblyopia; the red and green boxes represent FI in the viewing in LE in LEV and RE in REV in strabismic subject without amblyopia and control subject, respectively. Rightward and upward movements correspond to positive horizontal and vertical axes, respectively. Both the strabismus with and without amblyopia subjects have exotropia where the RE/AE is shifted to the right in LE/FEV and the LE/FE is shifted to the left in RE/AEV. AE = amblyopic eye; FE = fellow eye; LE = left eye; RE = right eye.

Previous research has established that FI in amblyopia and strabismus is associated with changes in normal FEMs and the presence of nystagmus, which consists of alternating quick phases and slower drift phases. Although BCEA summarizes the dispersion of fixation, it does not identify nystagmus, nor does it provide detailed metrics for distinguishing specific types of FEM abnormalities. To detect nystagmus, raw eye-position traces were examined (Fig 2). The Engbert and Kliegl algorithm was used to identify fixational saccades and nystagmus quick phases.⁵⁷ Slow drifts were defined as the intervals between these rapid movements. Because fast FEMs (fixational saccades and quick phases) follow a predictable saccadic main-sequence relationship,^58,59 all detected fast events were evaluated against this relationship to exclude artifacts or noisy detections.

Figure 2.

Time series and output of DBSCAN analysis (eye deviation) in control, strabismus without amblyopia, and strabismic amblyopia subjects. Time-series eye-position traces (top panels) and eye-deviation analyses (bottom panels) are shown for 3 representative subjects: (A) control, (B) strabismus without amblyopia, and (C) strabismic amblyopia, under fellow-eye/left-eye viewing (FEV/LEV; left column) and amblyopic-eye/right-eye viewing (AEV/REV; right column). Time-series eye positions (top panels): horizontal (solid lines) and vertical (dashed lines) eye positions are plotted over an 8-second segment of the fixation trial. Black traces represent the fellow/left eye, and blue traces represent the amblyopic/right eye. Positive values indicate rightward or upward movements. Fast fixation eye movements (FEMs) are marked with solid arrows (amplitudes in degrees) and slow drifts with dotted arrows (velocities in degrees/second). Numeric labels denote fast-phase amplitude and slow-phase drift velocity for each eye. (C) Fixation eye movements from a strabismic amblyopia subject diagnosed with fusion maldevelopment nystagmus as indicated by reversal in direction of nystagmus between FEV vs. AEV. The “=” symbol indicates a break in the y-axis to accommodate larger misalignments. Compared with the control subject (A), both the strabismus without amblyopia (B) and strabismic amblyopia (C) subjects exhibit larger fast-phase amplitudes and faster drift velocities in the viewing eye; however, the greatest interocular differences both within viewing conditions as well as across viewing conditions were noted in the amblyopic subject compared to the other 2 subjects. Eye-deviation and DBSCAN analysis (bottom panels), horizontal (thick lines) and vertical (thin lines) eye-deviation traces show the moment-to-moment binocular alignment. Density-Based Spatial Clustering and Application with Noise clustering identifies periods of similar alignment states: green clusters = well aligned, warm-colored clusters = misaligned. Two summary metrics are displayed: % time well aligned, the proportion of time spent within well-aligned states and time-based weighted mean deviation, which integrates both the magnitude of deviation and the duration spent in each cluster. The control participant shows a single well-aligned cluster (100% well aligned, lowest weighted mean deviation). Both strabismic subjects show only misaligned clusters (0% well aligned), with the strabismic amblyopia subject demonstrating the largest and most persistent deviations and greater interocular difference between AEV and FEV. DBSCAN = Density-Based Spatial Clustering and Application with Noise.

To further illustrate these metrics, Figure 2 (top panels) displays eye-position traces for the fellow/left and amblyopic/right eyes across viewing conditions for a control participant (A), a participant with strabismus without amblyopia (B), and a participant with strabismic amblyopia (C). The strabismic amblyopia subject exhibits fusion maldevelopment nystagmus, characterized by the hallmark reversal of quick-phase direction between the FEV and AEV conditions. In both viewing states, the slow phases remain directed nasally relative to the viewing eye: during FEV, the fellow (left) eye views and the slow phases drift rightward (nasally), whereas during AEV, when the amblyopic (right) eye views, the nystagmus reverses direction such that the slow phases again move nasally (leftward). This consistent nasalward drift across viewing conditions is a defining feature of fusion maldevelopment nystagmus. Fast-FEM amplitudes and slow-FEM velocities were lowest in the control subject and progressively larger in the nonamblyopic strabismic and amblyopic subjects. Also, the interocular difference across viewing condition in fast FEM amplitude and slow FEM velocity was greater in the strabismic amblyopic subject compared to the strabismus without amblyopia and control subject.

Quantification of Time-Based Control of Eye Deviation Using Density-Based Spatial Clustering and Application with Noise

To evaluate temporal control of eye alignment, the horizontal and vertical eye-position signals were smoothed using a moving average filter to remove rapid eye movements. The instantaneous difference between the 2 eyes was then used to compute a composite deviation signal (Equation [1.1]) and was calculated for each trial and subject.

A 2-dimensional Density-Based Spatial Clustering and Application with Noise clustering algorithm was applied to the deviation data to identify periods of stable alignment.60, 61, 62 Parameters were optimized to avoid splitting brief fluctuations into separate clusters. Clusters were classified as: well aligned, if their mean position lay within ±3.5° horizontally and ±2.0° vertically (thresholds based on the 95th percentile of control subjects’ variability), or misaligned, if the mean deviation exceeded these bounds.⁶⁰

To avoid unnecessary fragmentation, clusters with centroids within 1° horizontally and 0.5° vertically were merged. From the final clusters, 2 summary metrics were derived: (1) percentage of well aligned time (i.e., time without misalignment) and (2) time-weighted mean deviation, computed using Equations (1.2), (1.3), (1.4), reflecting the magnitude of eye deviation weighted by the duration of each cluster. Higher weighted means indicated more persistent and larger misalignment.²

$$\text{Composite}=\sqrt{\left({Horizontal}^2+{Vertical}^2\right)}$$ (1.1)

$$\begin{array}{l}F{T}_n=T_n/T_{total}\\ \left[T_{total}=T_1+T_2\right]\end{array}$$ (1.2)

$$ClusterComposit{e}_n=Composit{e}_n\ast F{T}_n$$ (1.3)

$$Mea{n}_{Weighted}=\left(\sum ClusterComposit{e}_n\right)/T_{total}$$ (1.4)

Figure 2 bottom panels in each box illustrate the eye deviation graphs representing the time-based control of eye deviation analysis, where thick lines represent horizontal eye deviation and thin lines represent vertical eye deviation, revealing increasing deviation from subjects A to C. The Density-Based Spatial Clustering and Application with Noise algorithm was applied with identified clusters shown in different colors; green colors represent well-aligned clusters, whereas warmer colors represent misaligned clusters. The control subject shows one well-aligned cluster (100% well aligned with the lowest weighted means values: 0.79 [LEV] and 0.24 [REV]), while both the strabismus without amblyopia and strabismic amblyopia subject show misaligned clusters (0% well aligned) with higher time-based weighted mean strabismus subject without amblyopia (LEV = 7.0, REV = 5.30) and strabismic subject with amblyopia (FEV = 19.3, AEV = 16.1). Furthermore, the difference in time-based weighted mean between AEV/REV and FEV/LEV was greatest in amblyopic subject and least in control subject. A complete summary of all FEM-related features and brief descriptions is provided in the Supplementary Material (eTables 1–3, available at www.ophthalmologyscience.org).

ML

Feature Engineering and Feature Selection

The comprehensive set of metrics described earlier served as an initial pool of 128 features for each ML classification task. Prior to feature selection, subjects with any missing feature values were identified and excluded from the data set to ensure data quality and integrity for model training. Any outliers, if present, were kept as they represented real clinical patient data. For feature selection, highly correlated features were marked for exclusion if the Pearson correlation coefficient between 2 features was >0.85. In each highly correlated feature pair, the feature that had the higher correlation to the predictor variable was kept, while the other feature with the lower correlation was excluded. An analysis of variance test was then performed on each feature stratified between the predictor variables; only features with a P value below a threshold value of 0.05 were retained for model training.

Model Creation & Hyperparameter Tuning

For each classification task, the categorical target variable (treated/mild strabismic amblyopia, moderate/severe strabismic amblyopia, strabismus without amblyopia, or control) was converted into numerical format using label encoding. After feature selection, the entire data set was split into training and testing sets using a 75% training and 25% testing split. Feature scaling using RobustScaler was applied to the training and testing sets to prevent features with larger numerical ranges to bias model training and slow convergence. The normalization parameters were obtained from the training data set and then applied to the testing set. Five distinct ML models were trained on each classification task and evaluated on performance: Logistic Regression (LR), Random Forest (RF) Classifier, Gaussian Naïve Bayes, Multilayer Perceptron Classifier, and Support Vector Machine Classifier. These selected models represent a multitude of common classification models across diverse algorithm families that include LR, RFs, Bayesian, neural networks, and support vector machines.⁶³ Hyperparameter tuning for each model was performed using GridSearchCV from Python’s sci-kit learn package with fivefold cross-validation on the training set. Each hyperparameter tuned model was then assessed with a series of performance metrics: accuracy, sensitivity, specificity, F₁ score, the area under the receiver operating characteristic curve (AUROC), and the area under the precision-recall curve (AUPRC). For each classification task, the best-performing model was selected based on a combination of the AUROC and the F₁ score. The confusion matrix for the best performing model was also shown. The standard deviation of each performance metric was estimated using 1000 bootstrapped samples from the testing data set.

Explainable AI with Shapley Additive exPlanations

To enhance the interpretability and provide transparent insights into each model’s decision-making process, the Shapley Additive exPlanations (SHAP) module was employed.⁶⁴ Shapley Additive exPlanations is a game-theoretic approach that explains the output of any ML model by computing the contribution of each feature to the prediction for a specific instance. Shapley Additive exPlanations values were calculated for the best performing model for each classification task. Shapley Additive exPlanations values quantify each feature’s impact on the model’s output probability for each class. In the case of binary classification models, 2 arrays of SHAP values were generated, one for each class. To simplify interpretation, only the SHAP values corresponding to the model’s predicted class were retained. The data sets used and analyzed during the present study are available from the corresponding author on reasonable request. The underlying code for this study and training/validation data sets is not publicly available but may be made available to qualified researchers on reasonable request to the corresponding and senior author (F.G.).

Results

A total of 120 strabismus subjects with and without amblyopia (mean age 16 ± 13.6, 68 women [56.7%]; 52 men [43.3%]) and 39 controls (mean age 14 ± 11.2, 26 women [66.7%]; 13 men [33.3%]) were included in the study. There were no differences in age controls: (14 ± 11 years; range, 4–64 years), strabismus subjects with and without amblyopia (16 ± 13 years; range, 3–69 years) (t test P = 0.49), and sex (χ², P = 0.27) between the 2 groups. Similarly, there was no significant difference in the distribution of race between strabismus with and without amblyopia (76 White [63.3%], 11 African American [9.2%], 11 Hispanic/Asian [9.2%], 22 multicultural [18.3%]) and controls (20 White [51.3%], 6 African American [15.4%], 8 Hispanic/Asian [20.5%], 5 multicultural [12.8%]) (χ², P = 0.14). Among the 73 participants with strabismic amblyopia, 22 had exotropia: 11 (15%) constant exotropia, 3 (4.1%) intermittent exotropia, 1 (1.4%) had infantile exotropia, and 7 (9.6%) consecutive exotropia and 51 had esotropia: 23 (31.5%) infantile esotropia, 18 (24.7%) partially accommodative esotropia, 8 (11%) fully accommodative esotropia, and 2 (2.7%) consecutive esotropia. Of 47 strabismus participants without amblyopia, 28 had exotropia, 7 (14.9%) constant exotropia, 15 (31.9%) intermittent exotropia, and 6 (12.8%) consecutive exotropia, and 19 had esotropia, including 14 (29.8%) infantile esotropia and 5 (10.6%) partially accommodative esotropia. A significant difference in the type of strabismus was observed between strabismus with and without amblyopia groups (Fisher exact test; P < 0.001). Strabismus without amblyopia group had greater number of intermittent exotropia participants than strabismus with amblyopia group. Control participants were allowed to have small, asymptomatic phorias within established normative limits to reflect physiological variability in the general population. In this cohort, 9 of 39 control participants exhibited near exophoria in the range of 2 to 4 prism diopters, which is considered within normal limits and substantially smaller than the exodeviation observed in participants with intermittent or constant exotropia.⁵⁰

Machine-learning models were developed and evaluated across 6 classification tasks using data from 4 subject groups: 24 individuals with treated or mild strabismic amblyopia, 43 individuals with moderate to severe strabismic amblyopia, 38 individuals with strabismus without amblyopia, and 37 healthy controls. Six classification tasks were evaluated:

Classification Task 1: moderate/severe strabismic amblyopia versus strabismus without amblyopia; Classification Task 2: strabismus without amblyopia versus controls; Classification Task 3: moderate/severe strabismic amblyopia versus controls; Classification Task 4: treated/mild strabismic amblyopia versus strabismus without amblyopia; Classification Task 5: treated/mild strabismic amblyopia versus controls; and Classification Task 6: strabismic amblyopia and strabismus without amblyopia versus controls.

The best performing models in classification tasks 1 to 3 were evaluated using data collected from additional participants over a 6-month period after the initial model development. This evaluation cohort included 6 subjects with moderate to severe strabismic amblyopia, 9 subjects with strabismus without amblyopia, and 2 control subjects. Due to the availability of only a single subject with mild strabismic amblyopia in the additional cohort, classification tasks 4 and 5 could not be evaluated using the newly collected data.

Classification Task 1: Moderate/Severe Strabismic Amblyopia versus Strabismus without Amblyopia

Fourteen features were selected for this classification task: 4 features were monocular, 3 were interocular ratios within viewing conditions, 7 were interocular ratios across viewing conditions, and none were binocular-coordination FEM metrics. Of the 7 within viewing condition features, all were from AEV with all monocular FEM features measuring the amblyopic viewing eye and the 3 interocular ratios capturing greater FEM abnormalities in the nonviewing FE. After hyperparameter tuning, performance metrics are summarized in Table 1. The RF model achieved the best overall performance (accuracy 0.81 ± 0.09, sensitivity 0.88 ± 0.12, specificity 0.77 ± 0.12, F₁ score 0.77 ± 0.12, AUROC 0.85 ± 0.09, AUPRC 0.77 ± 0.14; Fig 3A). Shapley Additive exPlanations analysis revealed the most influential feature as the interocular ratio of FI across viewing conditions (mean SHAP 0.010), followed by similar SHAP values for the other 8 most influential features (Fig 4A). The distribution of the top 9 features is available in eTable 4 (available at www.ophthalmologyscience.org).

Table 1.

Performance Metrics of Machine-Learning Models in Each Classification Task

Model	Accuracy (SD)	Sensitivity (SD)	Specificity (SD)	F1 Score (SD)	AUROC (SD)	AUPRC (SD)
Task 1: moderate/severe SA vs. S
LR	0.6699 (0.103)	0.6366 (0.177)	0.6900 (0.131)	0.5799 (0.149)	0.8073 (0.097)	0.7223 (0.151)
RF	0.8110 (0.086)	0.8777 (0.123)	0.7694 (0.117)	0.7705 (0.119)	0.8466 (0.090)	0.7682 (0.136)
GNB	0.5750 (0.106)	1.0000 (0.000)	0.3107 (0.127)	0.6332 (0.115)	0.7668 (0.105)	0.7017 (0.148)
MLP	0.8086 (0.086)	0.6244 (0.181)	0.9228 (0.078)	0.6957 (0.153)	0.9030 (0.080)	0.7971 (0.162)
SVM	0.6687 (0.106)	1.0000 (0.000)	0.4609 (0.140)	0.6903 (0.117)	0.8191 (0.097)	0.7086 (0.162)
Task 2: S vs. control
LR	0.9449 (0.052)	0.8506 (0.145)	1.0000 (0.000)	0.9118 (0.097)	0.8880 (0.111)	0.9159 (0.092)
RF	0.8969 (0.071)	0.8606 (0.133)	0.9187 (0.081)	0.8525 (0.111)	0.9192 (0.074)	0.9178 (0.082)
GNB	0.8400 (0.083)	0.5684 (0.194)	1.0000 (0.000)	0.7041 (0.171)	0.9882 (0.020)	0.9826 (0.032)
MLP	0.7911 (0.093)	0.5803 (0.190)	0.9166 (0.081)	0.6542 (0.169)	0.8412 (0.107)	0.8412 (0.111)
SVM	0.8437 (0.082)	0.5738 (0.194)	1.0000 (0.000)	0.7087 (0.169)	0.9520 (0.057)	0.9481 (0.065)
Task 3: moderate/severe SA vs. control
LR	0.8576 (0.080)	0.6812 (0.166)	1.0000 (0.000)	0.7981 (0.125)	1.0000 (0.000)	1.0000 (0.000)
RF	0.7975 (0.088)	0.5563 (0.168)	1.0000 (0.000)	0.6992 (0.147)	0.9705 (0.034)	0.9659 (0.042)
GNB	0.7497 (0.098)	0.4469 (0.169)	1.0000 (0.000)	0.5983 (0.169)	0.9812 (0.023)	0.9813 (0.025)
MLP	0.8979 (0.066)	0.7730 (0.139)	1.0000 (0.000)	0.8647 (0.094)	0.9698 (0.032)	0.9673 (0.038)
SVM	0.6962 (0.103)	0.3344 (0.167)	1.0000 (0.000)	0.4773 (0.193)	0.9796 (0.026)	0.9772 (0.030)
Task 4: treated/mild SA vs. S
LR	0.6959 (0.117)	0.8015 (0.130)	0.5222 (0.230)	0.7573 (0.109)	0.5995 (0.185)	0.7210 (0.150)
RF	0.5618 (0.125)	0.6985 (0.150)	0.3311 (0.214)	0.6563 (0.123)	0.5494 (0.167)	0.7482 (0.123)
GNB	0.4991 (0.123)	0.2965 (0.145)	0.8338 (0.161)	0.4063 (0.168)	0.6510 (0.181)	0.7392 (0.147)
MLP	0.6251 (0.125)	0.7967 (0.126)	0.3373 (0.209)	0.7176 (0.113)	0.5159 (0.177)	0.6900 (0.148)
SVM	0.6249 (0.123)	1.0000 (0.000)	0.0000 (0.000)	0.7618 (0.098)	0.3281 (0.168)	0.5804 (0.142)
Task 5: treated/mild SA vs. control
LR	0.6896 (0.116)	0.5070 (0.219)	0.8015 (0.130)	0.5263 (0.192)	0.7697 (0.154)	0.7716 (0.154)
RF	0.7463 (0.112)	0.6536 (0.226)	0.8018 (0.132)	0.6303 (0.187)	0.8151 (0.122)	0.7590 (0.171)
GNB	0.9389 (0.060)	0.8336 (0.162)	1.0000 (0.000)	0.8996 (0.111)	0.8660 (0.133)	0.9032 (0.108)
MLP	0.6873 (0.114)	0.3286 (0.204)	0.9021 (0.093)	0.4099 (0.220)	0.8340 (0.109)	0.7944 (0.146)
SVM	0.7469 (0.111)	0.3254 (0.205)	1.0000 (0.000)	0.4554 (0.236)	0.8525 (0.112)	0.8594 (0.116)
Task 6: SA + S vs. control
LR	0.8915 (0.052)	0.9589 (0.040)	0.7568 (0.129)	0.9204 (0.040)	0.9314 (0.044)	0.9677 (0.022)
RF	0.8632 (0.057)	0.9609 (0.039)	0.6698 (0.141)	0.9016 (0.044)	0.9079 (0.056)	0.9523 (0.034)
GNB	0.7749 (0.069)	0.7036 (0.096)	0.9183 (0.081)	0.8029 (0.069)	0.9347 (0.041)	0.9680 (0.023)
MLP	0.8601 (0.058)	0.8730 (0.068)	0.8344 (0.110)	0.8906 (0.049)	0.9089 (0.053)	0.9597 (0.027)
SVM	0.8368 (0.061)	0.9605 (0.040)	0.5869 (0.149)	0.8852 (0.047)	0.9421 (0.036)	0.9736 (0.019)

AUPRC = area under precision-recall curve; AUROC = area under receiver operating characteristic curve; C = control; GNB = Gaussian naïve bayes; LR = logistic regression; MLP = multilayer perceptron classifier; RF = random forest; S = strabismus without amblyopia; SA = strabismic amblyopia; SD = standard deviation; SVM = support vector machine.

Figure 3.

Model diagnostics for classification tasks with moderate/severe strabismic amblyopia. Receiver operating characteristic (ROC) curves with the area under the curve (AUROC) value and the confusion matrix are shown for the best model in (A) Classification Task 1, (B) Classification Task 2, and (C) Classification Task 3. Accuracy for the best models are 81%, 89%, and 90%, respectively. S = strabismus without amblyopia; SA = strabismic amblyopia; SD = standard deviation.

Figure 4.

Shapley Additive exPlanations analysis (SHAP) for classification tasks with moderate/severe strabismic amblyopia. Shapley Additive exPlanations plots were generated for the best performing model in (A) Classification Task 1, (B) Classification Task 2, and (C) Classification Task 3. Each plot graphs the global feature importance and displays the mean absolute SHAP value for the top 9 features of the predicted class in each subject. The mean SHAP value quantifies the average impact of each feature on the model’s output across all subjects. The viewing condition and eye was designated fellow-eye and amblyopic-eye viewing (FEV/AEV) for amblyopic subjects and left-eye and right-eye viewing (LEV/REV) for strabismic subjects without amblyopia and controls. AE = amblyopic eye; C = control; FE = fellow eye; FEM = fixation eye movement; IOD = interocular difference; IOR = interocular ratio; LE = left eye; RE = right eye; S = strabismus without amblyopia; SA = strabismic amblyopia. ∗Although these 2 are matching features, the first one is the 25th percentile feature, while the second one is the 75th percentile feature. ^wWithin the same viewing conditions. ^aAcross viewing conditions.

Classification Task 2: Strabismus without Amblyopia versus Controls

Thirty-three features were selected for this classification task: 11 features were monocular, 18 were interocular ratios within viewing conditions, 0 were interocular ratios across viewing conditions, and 4 were binocular-coordination FEM metrics. The RF model achieved the best overall performance (accuracy 0.90 ± 0.07, sensitivity 0.86 ± 0.13, specificity 0.92 ± 0.08, F₁ score 0.85 ± 0.11, AUROC 0.92 ± 0.07, AUPRC 0.92 ± 0.08; Fig 3B). Shapley Additive exPlanations analysis revealed the most influential feature as the time-based weighted mean measured in FEV (mean SHAP 0.08), followed by the percent time well aligned measured in FEV (mean SHAP 0.06) and AEV (mean SHAP 0.03); the remaining 6 features had similar SHAP values (Fig 4B). The distribution of the top 9 features is available in eTable 5 (available at www.ophthalmologyscience.org).

Classification Task 3: Moderate/Severe Strabismic Amblyopia versus Controls

Twenty-nine features were selected for this classification task: 9 features were monocular, 7 were interocular ratios within viewing conditions, 7 were interocular ratios across viewing conditions, and 6 were binocular-coordination FEM metrics. The Multilayer Perceptron Classifier model achieved the best overall performance (accuracy 0.90 ± 0.07, sensitivity 0.77 ± 0.14, specificity 1.00 ± 0.00, F₁ score 0.86 ± 0.10, AUROC 0.97 ± 0.03, AUPRC 0.97 ± 0.04; Fig 3C). Shapley Additive exPlanations analysis revealed the most influential feature as the interocular ratio of the AE over the FE for the slow FEM velocity during FEV (mean SHAP 0.09), followed by the interocular ratio of the slow over fast FEM velocity of the AE in FEV (mean SHAP 0.07; Fig 4C). The distribution of the top 9 features is available in eTable 6 (available at www.ophthalmologyscience.org).

Classification Task 4: Mild/Treated Strabismic Amblyopia versus Strabismus without Amblyopia

Only 2 features were selected for this classification task, one was an interocular ratio within viewing conditions FEM feature, while the other was an interocular ratio across viewing conditions FEM feature. The LR model achieved the best overall performance (accuracy 0.70 ± 0.12, sensitivity 0.80 ± 0.13, specificity 0.52 ± 0.23, F₁ score 0.75 ± 0.11, AUROC 0.60 ± 0.19, AUPRC 0.72 ± 0.15; Fig 5A). The distribution of the 2 features is available in eTable 7 (available at www.ophthalmologyscience.org).

Figure 5.

Model diagnostics for classification tasks with treated/mild strabismic amblyopia and all amblyopia and strabismus combined. Receiver operating characteristic (ROC) curves with the area under the curve (AUROC) value, confusion matrix, and SHAP plots are shown for the best model in (A) Classification Task 4, (B) Classification Task 5, and (C) Classification Task 6. Accuracy for the best model in classification task 4 is 69%, classification task 5 is 94%, and classification task 6 is 89%. S = strabismus without amblyopia; SA = strabismic amblyopia; SD = standard deviation; SHAP = Shapley Additive exPlanations.

Classification Task 5: Mild/Treated Strabismic Amblyopia versus Controls

Twenty-seven features were selected for this classification task: 11 features were monocular, 8 were interocular ratios within viewing conditions, 2 were interocular ratios across viewing conditions, and 6 were binocular-coordination FEM metrics. The Gaussian Naïve Bayes model achieved the best overall performance (accuracy 0.94 ± 0.06, sensitivity 0.83 ± 0.16, specificity 1.00 ± 0.00, F₁ score 0.90 ± 0.11, AUROC 0.87 ± 0.13, AUPRC 0.90 ± 0.11; Fig 5B). Shapley Additive exPlanations analysis revealed the most influential feature as the time-based weighted mean measured in FEV (mean SHAP 0.14), followed by the interocular ratio of FI across viewing conditions (mean SHAP 0.10) and percent time well aligned measured in FEV (mean SHAP 0.09); the remaining 6 features had similar SHAP values. The distribution of the top 9 features is available in eTable 8 (available at www.ophthalmologyscience.org).

Classification Task 6: Strabismic Amblyopia and Strabismus without Amblyopia versus Controls

Beyond discrimination among specific clinical subgroups, we evaluated classification performance in a clinically relevant screening framework in which all amblyopia and strabismus cases were grouped together against controls. In this combined classification task, model performance remained strong, achieving high accuracy with the LR model (accuracy 0.89 ± 0.05, sensitivity 0.96 ± 0.04, specificity 0.76 ± 0.13, F₁ score 0.92 ± 0.04, AUROC 0.93 ± 0.04, AUPRC 0.97 ± 0.02; Fig 5C). These results indicate that fixation-related eye movement features provide robust separation between affected children and controls, even when heterogeneous phenotypes are combined.

Model Performance on Newly Collected Participants

Performance on the additional evaluation cohort using the best performing model in classification tasks 1 to 3 yielded an accuracy of 87%, 90%, and 88%, respectively, as shown in Figure 6. The same scaler that was used for feature scaling during development of the original models was applied to this prospective data set.

Figure 6.

Model diagnostics for classification tasks with newly collected participants. Confusion matrix are shown for the best model in each respective classification task, (A) Classification Task 1, (B) Classification Task 2, and (C) Classification Task 3 on prospectively collected data set. Accuracy for the best model in each classification task is 87%, 90%, and 88%, respectively. Machine-learning models using FEM features accurately distinguished strabismic amblyopia, strabismus, and controls, with strongest performance in more severe disease. Results support FEM-based ML as a future scalable screening tool for clinical differentiation. FEM = fixation eye movement; ML = machine learning; S = strabismus without amblyopia; SA = strabismic amblyopia.

Discussion

This study describes the development and evaluation of different ML models across 6 binary classification tasks: moderate/severe strabismic amblyopia versus strabismus without amblyopia (Classification Task 1), strabismus without amblyopia versus controls (Classification Task 2), moderate/severe strabismic amblyopia versus controls (Classification Task 3), the more challenging distinctions involving treated/mild strabismic amblyopia versus strabismus without amblyopia (Classification Task 4), and treated/mild strabismic amblyopia versus controls (Classification Task 5). All amblyopia and strabismus cases were also classified against controls in Classification Task 6 to demonstrate feasibility of a clinically relevant screening framework.

The best models within classification tasks 1 to 3, which focused on moderate/severe forms of strabismic amblyopia and strabismus, demonstrated strong generalizability on an independent test set, achieving accuracies of 87%, 90%, and 88%, respectively. Shapley Additive exPlanations analysis revealed 3 consistent patterns across these classification tasks: (1) monocular and interocular ratios across viewing conditions FEM features, which reflected abnormalities in the AE compared to the FE, were the strongest indicators of amblyopia; (2) monocular and interocular within viewing conditions FEM features, which reflected abnormalities in the nonviewing eye were most predictive of strabismus without amblyopia; and (3) binocular-coordination metrics—particularly deviation magnitude—were key markers of strabismus regardless of amblyopia status. Importantly, interocular ratios calculated within and across viewing conditions strengthened model performance by normalizing FEM differences within each subject, reducing confounding influences such as age, attention, and cooperation.^28,65

Classification task‑specific patterns highlight how different FEM features capture distinct aspects of oculomotor pathology. In classification task 1, which differentiates moderate/severe amblyopia from strabismus without amblyopia, interocular FI across viewing conditions emerged as the most influential feature. This metric reflects fixation stability in the viewing eye under each monocular condition—comparing the AE during AEV with the FE during FEV—thereby providing a direct measure of how fixation changes when visual feedback is available. In the amblyopic group, greater interocular FI was closely linked to higher interocular fast-FEM velocity, largely driven by increased fast-FEM amplitudes in the AE. These abnormalities were accompanied by fewer microsaccades (<1°) and an increase in larger fixational saccades (>1°), a pattern consistent with prior reports in moderate/severe amblyopia.^{26,27,29,34,66} By contrast, nonviewing eye FEM metrics were similar in amblyopic and nonamblyopic strabismus, limiting their value for discrimination. The model’s integration of both microsaccade and fixational-saccade frequencies—characterized by reduced microsaccade frequencies and increased fixational-saccade frequencies in AEV compared with FEV in the strabismic amblyopia group, and minimal differences between REV and LEV in the strabismus without amblyopia group—further highlighted FEM patterns that reliably distinguish these 2 groups.

Classification tasks 2 and 3 expanded on these findings. The best model in classification task 2, which compared strabismus without amblyopia to controls, was dominated by binocular-coordination metrics and nonviewing-eye abnormalities characteristic of strabismus. The best model in classification task 3, which distinguished moderate/severe strabismic amblyopia from controls, incorporated the most balanced mix of all 4 FEM feature classes. Its top features included interocular ratios reflecting greater slow-FEM amplitude and velocity abnormalities in the nonviewing eye, along with binocular-coordination metrics and amblyopic-eye FEM deficits observed in AEV condition. This diverse feature set likely reflects the combined amblyopic and strabismic pathology in this group and explains why the best model in classification task 3 performed best among the 3.

Classification tasks 4 and 5 evaluated distinctions involving treated or mild amblyopia. Classification task 4, which compared treated/mild amblyopia to strabismus without amblyopia, performed poorly with the best model selecting only 2 features. Its AUROC of 0.58 highlights the difficulty of distinguishing treated/mild amblyopia from strabismus without amblyopia when both groups share similar strabismic characteristics and FEM abnormalities are subtle. In contrast, classification task 5—which compared treated/mild amblyopia to controls—performed much better, with the best model selecting features from all 4 FEM categories and identifying binocular-coordination metrics as the top predictor. Across both classification tasks, each model’s consistent selection of interocular ratios within and across viewing conditions underscores their sensitivity for detecting small within-subject FEM asymmetries, making them more informative than monocular FEM metrics alone in identifying treated/mild strabismic amblyopia features. Finally, the strong performance of classification task 6, which evaluated the combined grouping of all prior amblyopia and strabismus cases against controls, demonstrates the feasibility of a screening-oriented task in differentiating any type of pathology from controls. Collectively, these results illustrate the complementary roles of FEM abnormalities observed in the viewing and nonviewing eye, binocular coordination metrics, and interocular comparisons in characterizing the diverse oculomotor signatures of strabismus and amblyopia across the severity spectrum.

These findings can be interpreted within the broader framework of established mechanisms underlying FI in amblyopia and strabismus. Under normal conditions, fixation is most stable during binocular viewing, when the brain combines visual inputs from both eyes to maintain accurate gaze.67, 68, 69, 70 Monocular viewing reduces visual feedback and weakens the neural integrator responsible for transforming premotor velocity signals into sustained eye-position commands, resulting in increased drift amplitude and velocity, even in the FE.^20,26,71,72 In amblyopia, degraded visual input could impair the function of the neural integrator, producing disproportionately greater drift and reduced fixation stability when the AE is the viewing eye. Prior studies describe increased instability in amblyopia involving both drifts and fixational saccades, although whether such instability is a cause or consequence of reduced visual acuity remains unresolved.^26,28,29,34 Because the neural integrator receives visual input from each eye separately, loss of visual feedback reduces gaze stability in both eyes but most prominently in the eye with poorer visual input. Consistent with this physiology, our data show increased horizontal drift in both eyes of amblyopic participants, with the greatest instability occurring during AEV.

Additional mechanistic insight comes from work in strabismus.^37,73, 74, 75, 76, 77 In primate models, FI has been linked to abnormal development of vergence pathways. Stimulation of the rostral superior colliculus (SC) can shift the strabismic angle, suggesting that SC-mediated influences on vergence neurons contribute to misalignment. Physiological and computational studies also implicate SC-modulated interactions between omnipause neurons and long-lead burst neurons in the generation of microsaccades and saccades. Together, these findings suggest that fixation scatter in strabismus reflects both altered fixational saccade generation and SC-linked mechanisms that maintain abnormal ocular alignment. Importantly, these deficits occur even when visual acuity is normal, indicating that long-standing decorrelation of binocular input disrupts the development and function of fusion mechanisms, leading to persistent FI.

Taken together, these physiological principles provide a coherent explanation for the FEM patterns observed across our classification tasks. Loss of binocular cues—whether from amblyopic visual degradation or strabismic decorrelation—reduces gaze stability in both eyes, but the effect is greatest in the AE. The distinct instability patterns with and without visual feedback, that is, viewing versus nonviewing eye, align with this framework: nonviewing eye FEM abnormalities increased similarly in both strabismus-only and strabismic-amblyopia groups due to reduced visual feedback in conjunction with the eye deviation, whereas viewing eye had better fixation stability in strabismus without amblyopia but remained impaired in amblyopia, where visual feedback is intrinsically degraded. These physiological distinctions mirror the SHAP feature rankings observed across the various classification tasks: nonviewing eye FEM metrics best predicted strabismus, binocular-coordination measures captured positional disparities in the strabismic groups, and viewing-eye metrics were most informative for amblyopia classification. Collectively, these insights reinforce the biological plausibility of the FEM-based classifiers and support the task-specific distinctions observed in our analyses.

This study has several limitations. First, the modest sample size constrained granular classification of strabismus with and without amblyopia; for example, certain strabismus phenotypes such as well-controlled intermittent exotropia can exhibit visual acuity and stereoacuity comparable to controls, making classification even more challenging.

Furthermore, stratification by coexisting nystagmus, which is more prevalent in strabismic amblyopia, could not be assessed in this study, although our prior work showed strong deep-learning performance in amblyopic subjects both with and without nystagmus, likely because even those without nystagmus exhibit alterations in physiologic FEMs.⁴¹ We partially mitigated small-sample effects by validating on an independent test set, using standardized acquisition and preprocessing, and leveraging interocular ratios (within and across conditions) to normalize subject-level variability, but larger, phenotypically stratified data sets remain necessary. The weaker performance observed in classification task 4 further highlights the challenge of detecting subtle abnormalities in mild/treated amblyopia, where strabismic instability can overshadow small effects. Additional binocular-viewing recordings, higher-resolution FEM features, and longer recordings may increase sensitivity to mild deficits.

Second, several subject-level factors warrant consideration. Uncorrected refractive error can influence FEMs through visual blur, degrading foveal image quality and modestly increasing retinal error. However, these effects are generally smaller in magnitude and mechanistically distinct from the abnormalities observed in amblyopia, which are characterized by disruption of physiologic FEMs and prominent interocular asymmetries that scale with disease severity and likely reflect impaired sensory–motor integration.^26,78,79 Consistent with prior work showing that visual acuity can influence fixation metrics,²⁸ uncorrected refractive error may introduce modest variability into fixation-based measures; importantly, both the magnitude and spatial profile of these changes differ from those observed in amblyopia or strabismus. Because participants were tested under best-corrected viewing conditions, the potential confounding effects of refractive error were minimized. Future studies should evaluate model performance under both corrected and uncorrected viewing conditions to better simulate clinical screening environments. Although participants spanned a broad age range (3–69 years), prior work suggests that interocular fixation metrics which are abnormal in amblyopia and strabismus are relatively stable across the lifespan in visually normal individuals, and age distributions were comparable between our study groups.⁶⁵ Furthermore, amblyopia-related abnormalities—such as increased FI and altered FEM dynamics—are well established to be robust, severity-dependent, and largely independent of age. Additionally, the relatively short testing duration (approximately 3–5 minutes total, including calibration and brief breaks in between trials) and standardized, pediatric-friendly recording conditions further reduce variability related to attention, fatigue, or task compliance across age groups. Prior studies of FEMs in amblyopia have used fixation epochs ranging from 6 to 45 seconds and have consistently demonstrated measurable abnormalities.^18,30,34,53 Future work will evaluate whether shorter recording durations are sufficient to support reliable classification, with the goal of further optimizing feasibility for screening applications. Together, these considerations suggest that neither refractive error nor age substantially influenced the observed group differences, although both warrant prospective evaluation in larger cohorts. Importantly, the distinct magnitude and interocular structure of amblyopia and strabismus-related fixation abnormalities may confer resilience to modest variability introduced by refractive blur or age in screening applications.

Finally, scalability and external generalizability remain important challenges. High-quality eye movement recordings currently require specialized hardware, technical expertise, and standardized acquisition procedures. In addition, our data were collected at a single center and may not capture variability arising from differences in equipment, calibration methods, or clinical populations. We partially addressed these concerns through standardized preprocessing, validation on an independent test set, and the use of interocular ratio metrics that normalize subject-level variability. Nevertheless, multicenter collaborations, shared data sets, and automated analysis pipelines will be essential for broader clinical implementation. Although SHAP analysis improves model interpretability, it does not establish causality, and some predictive features may not directly reflect underlying disease mechanisms. Future behavioral, physiologic, and multicenter validation studies will therefore be necessary to confirm the biological significance and clinical robustness of these FEM-based classifiers.

In summary, this study presents the first ML framework designed to distinguish strabismus with and without amblyopia using FEMs and builds upon a literature of potential ML approaches to enhance the diagnosis and management of strabismus and amblyopia. We show that simple, prebuilt, and widely available binary classification models are sufficient to achieve high discrimination in predicting these 2 similarly presenting eye pathologies. More importantly, this study builds upon prior work that measures FEM patterns of viewing and nonviewing eye are critical to differentiate strabismus with and without amblyopia. These findings lay the foundation for future multicenter efforts to collect FEM data across the full spectrum of strabismic and amblyopic pathologies and to develop multiclass models capable of differentiating various disease phenotypes, thereby enhancing the clinical utility of ML approaches in the management of common childhood eye diseases.

Manuscript no. XOPS-D-26-00119.