A Digital Twin–Inspired Patient-Specific Eye–Lens Simulation Framework for Pediatric Postcycloplegic Spectacle Prescription: A Prospective Agreement Study

Abstract

Purpose

To evaluate the agreement of a digital twin–inspired patient-specific eye–lens simulation framework and reference-standard subjective refraction for pediatric postcycloplegic spectacle prescription.

Methods

In this prospective randomized testing sequence agreement study, 183 right eyes from 183 children (mean age, 9.9 ± 2.0 years) underwent traditional autorefraction, artificial intelligence (AI) wavefront refraction, and masked subjective refraction. The framework used microelectromechanical systems (MEMS) Shack–Hartmann wavefront data, image quality gating, virtual trial-lens simulation, and perturbation-aware optimization to generate an AI-based prescription estimate. Agreement was assessed for spherical equivalent (SE), Jackson cross-cylinder component at axis 0° (J0), and Jackson cross-cylinder component at axis 45° (J45).

Results

Traditional autorefraction differed significantly from subjective refraction in SE, J0, and J45: 0.21 ± 0.52 diopters (D) for SE (P < 0.001), −0.04 ± 0.21 D for J0 (P = 0.006), and 0.02 ± 0.13 D for J45 (P = 0.037). AI wavefront refraction showed smaller mean differences without statistically significant differences: 0.04 ± 0.35 D for SE (P = 0.153), −0.01 ± 0.19 D for J0 (P = 0.681), and 0.00 ± 0.15 D for J45 (P = 0.725). For SE, AI wavefront refraction showed narrower 95% limits of agreement than traditional autorefraction (−0.67 to 0.74 D vs. −0.81 to 1.22 D) and higher proportions within ≤0.25 D and ≤0.50 D (66.1% vs. 54.6%; 91.3% vs. 86.9%).

Conclusions

The digital twin–inspired patient-specific eye–lens simulation framework generated pediatric postcycloplegic prescription estimates that agreed more closely with subjective refraction than traditional autorefraction, particularly for SE. These findings support its feasibility as an upstream decision support component, although real-world spectacle wear feedback and longitudinal model updating require further study.

Translational Relevance

This study provides prospective clinical validation for an upstream component of a future human-in-the-loop ophthalmic digital twin for spectacle prescription by translating patient-specific wavefront data into virtual trial-lens simulation and AI-assisted refraction decision support.

Introduction

Spectacle prescription in children requires accurate estimation of refractive status while accounting for the child's cooperation, accommodative status, and visual responses during clinical refraction. Conventional prescribing strategies are largely based on objective measurements followed by subjective refinement, but the relationship between measured ocular optics and the final accepted prescription may vary across individuals. With the increasing availability of ocular biometry and wavefront aberrometry, there is growing interest in computational approaches that incorporate individual optical characteristics into refractive assessment.¹ Digital twin frameworks have recently been introduced into medicine and ophthalmology to support patient-specific virtual modeling, simulation, and decision support.^2–4 In the strict sense proposed by the National Academies of Sciences, Engineering, and Medicine, a complete digital twin should mimic the structure, context, and behavior of a physical system; be dynamically updated with data from its physical counterpart; have predictive capability; and inform decisions that create value. This distinction is relevant because a static digital model or a unidirectional digital shadow does not by itself constitute a complete digital twin. Recent human digital twin taxonomies further separate such models from one-time decision support models and closed-loop digital twins.^5,6 Under this framework, a complete ophthalmic digital twin for spectacle prescription would require not only a patient-specific virtual eye–lens model but also postdispensing feedback, longitudinal updating, and a bidirectional decision loop. The present study therefore uses the term digital twin–inspired in a restricted and taxonomy-aware sense: it evaluates an upstream patient-specific optical modeling and simulation component rather than a complete closed-loop ophthalmic digital twin. Thus, the digital twin–inspired designation refers to this patient-specific modeling and simulation sequence rather than to wavefront sensing alone.

In pediatric refraction, subjective refraction remains the clinical reference standard for spectacle prescribing because it incorporates both optical correction and perceptual feedback. However, subjective refinement can be challenging in children because attention span, understanding of task instructions, and fatigue may affect the reliability of responses.⁷ Traditional autorefraction is widely used as an efficient objective starting point, but its measurements may be influenced by assumptions embedded in static schematic eye models and by physiologic factors such as pupil size variation, accommodative microfluctuation, and tear film instability.^8,9 These issues are particularly relevant in pediatric postcycloplegic practice. Cycloplegia is commonly used to reveal refractive status in children, whereas final prescription decisions are often made during postcycloplegic reexamination, when accommodative tone may be recovering.^10,11 This creates a practical need for objective methods that can better approximate the subjective prescription endpoint in this clinical setting.

Wavefront aberrometry provides a more detailed description of whole-eye optical aberrations than conventional autorefraction and has been explored as a basis for objective prescription estimation.¹² Earlier correction strategies often focused on reducing or compensating higher-order aberrations. However, a prescription optimized for a single static optical state may be sensitive to changes in lens centration, pupil size, visual axis alignment, and other fitting-related factors during spectacle wear.¹³ Therefore, objective refraction algorithms may need to consider not only retinal image quality under idealized conditions but also the stability of optical performance under small perturbations. Image quality metrics, including the visual Strehl ratio, have been used to link wavefront-derived optical information with visual performance and subjective refraction endpoints.^14,15 More recently, machine learning methods have been applied to wavefront and ocular feature data to predict subjective refraction, suggesting a possible role for data-driven approaches in objective to subjective refraction translation.^8,16

Despite these advances, several gaps remain. First, many objective or artificial intelligence (AI)–based refraction methods still follow a measurement to prescription estimate paradigm and do not explicitly construct a patient-specific virtual eye–lens model in which candidate prescriptions can be simulated before clinical use. In this respect, they remain closer to objective measurement or one-time data-driven estimation than to an intervention simulation framework. In particular, few approaches describe how the physical trial-lens process of subjective refraction can be partially transferred into a virtual trial-lens simulation environment or how clinically relevant perturbations, such as pupil size variation, lens decentration, vertex distance variation, and fixation instability, are handled computationally. Accordingly, the methodological novelty of the present study is not another agreement comparison alone but the prospective evaluation of a simulation-based objective to subjective prescription pathway that combines patient-specific eye–lens modeling, in silico trial-lens testing, perturbation-aware optimization, and AI-based mapping to a subjective refraction–derived endpoint. Second, evidence remains limited in pediatric postcycloplegic reexamination, a setting in which accommodation recovery and child cooperation may both influence the final prescription. Third, it remains unclear whether a wavefront-based patient-specific simulation framework can provide prescription estimates that agree more closely with reference-standard subjective refraction than traditional autorefraction, particularly when refractive outcomes are evaluated using spherical equivalent (SE), Jackson cross-cylinder component at axis 0° (J0), and Jackson cross-cylinder component at axis 45° (J45) vector components.

The present study therefore developed and prospectively evaluated a digital twin–inspired patient-specific eye–lens simulation framework for pediatric postcycloplegic spectacle prescription. The framework was designed as the upstream modeling and simulation layer of a future closed-loop ophthalmic digital twin, rather than as a complete digital twin system. It incorporated image quality–controlled wavefront acquisition, patient-specific virtual optical eye model construction, virtual trial-lens simulation, perturbation-aware robust optimization, and AI-based mapping to subjective refraction–derived prescription endpoints. We conducted a prospective randomized testing sequence, within-subject agreement study to compare the AI-generated prescription estimates and traditional autorefraction against masked subjective refraction in 183 right eyes from 183 pediatric participants. Agreement was assessed using power vector components, paired differences, Pearson correlation, Bland–Altman limits of agreement, and absolute SE error thresholds.

Methods

Study Design and Participants

This study was designed as a prospective randomized testing sequence, within-subject agreement study. The study protocol was approved by the Ethics Committee of Beijing Tongren Hospital, Capital Medical University (approval number: TREC2024-KY131) and was conducted in accordance with the principles of the Declaration of Helsinki. The legal guardians of all pediatric participants were informed of the study objectives and procedures and provided written informed consent.

School-aged children who presented for refractive error correction or myopia management were recruited. The inclusion criteria were age between 6 and 18 years and the ability to cooperate with both wavefront aberrometry and subjective refraction. The main exclusion criteria included a history of ocular surgery, strabismus, amblyopia, pathologic corneal abnormalities such as keratoconus, systemic diseases affecting accommodation, and current or previous orthokeratology lens wear.

For statistical analysis, only the right eye of each participant was included. This prespecified eye-level analytic approach was used to avoid intereye correlation in paired agreement analyses.

Sample Size Rationale

The sample size estimation was based on a preliminary dataset of approximately 26 patients evaluated with the same digital twin–guided AI wavefront refraction algorithm. In the preliminary dataset, for SE, the mean bias of digital twin–guided AI wavefront refraction relative to clinical reference-standard subjective refraction was 0.01 ± 0.48 diopters (D), with 95% limits of agreement from −0.94 to 0.96 D. Because the present study was designed as an agreement study, the formal sample size was calculated to ensure adequate precision of the Bland–Altman limits of agreement (LoA) for SE. Assuming normally distributed paired SE differences, the sample size was calculated using the Bland–Altman precision formula shown above. Using the pilot standard deviation of paired SE differences of 0.48 D and setting the target half-width at 0.125 D, approximately half of a 0.25-D clinical refraction step, the required sample size was 169.9, corresponding to at least 170 analyzable eyes. The final cohort included 183 analyzable right eyes, exceeding this requirement and providing an expected SE LoA confidence half-width of approximately 0.12 D.

Clinical Workflow and Randomized Testing Sequence Design

All testing was performed during the postcycloplegic reexamination phase, after the effect of the cycloplegic agent had subsided and physiologic accommodation had recovered. This setting was intended to reflect the clinical stage at which spectacle prescriptions are commonly determined for children after cycloplegic assessment.

To reduce potential fatigue bias from repeated examinations and to control for order effects, participants were assigned by a computer-generated random number table to one of two testing sequences:

• Sequence A: objective wavefront aberrometry with AI prescription estimation, followed by standard autorefraction and subjective refraction

• Sequence B: standard autorefraction and subjective refraction, followed by objective wavefront aberrometry with AI prescription estimation

Standard autorefraction was performed using a KR-1 autorefractor (Topcon, Tokyo, Japan). Three consecutive measurements were automatically obtained and averaged according to the device workflow.

Subjective refraction was independently performed by the same experienced optometrist using a standard phoropter and visual acuity chart. The optometrist was masked to the AI refraction results. The endpoint for subjective refraction was defined as maximum plus to best visual acuity.

Digital Twin–Inspired Patient-Specific Eye–Lens Simulation Framework

Model Classification and Design Rationale

The framework was classified a priori as a digital twin–inspired personalized eye–lens simulation framework rather than as a complete human digital twin. The physical system was defined as the individual eye in a spectacle-prescribing context, and the virtual representation was defined as a patient-specific optical eye model reconstructed from quality-controlled wavefront measurements. This virtual model was coupled with a digital trial-lens library to simulate the optical consequences of candidate sphero-cylindrical prescriptions before generating a decision support output. Because the present implementation did not include longitudinal real-world spectacle wear feedback, continuous physiologic updating, or bidirectional closed-loop model revision, it is more accurately described as a personalized digital model used once for refraction decision support. Its digital twin–inspired elements include patient-specific data acquisition, virtual optical representation, simulation of candidate interventions, perturbation-aware robustness testing, AI-based prescription estimation, and prospective validation against masked subjective refraction. Accordingly, the present framework should be understood as a single-encounter, patient-specific optical simulation model inspired by digital twin principles, rather than as a dynamic physiologic digital twin with continuous feedback or adaptive updating. For clarity, the steps constituting the digital twin–inspired component were physical to virtual data mapping through image quality–controlled wavefront acquisition, patient-specific virtual optical eye model construction, virtual trial-lens simulation of candidate sphero-cylindrical prescriptions, perturbation-aware robust prescription optimization, and AI-based decision support output validated against masked subjective refraction. The overall study workflow and digital twin–inspired AI prescription pipeline are shown in Figure 1.

Figure 1.

Study workflow and digital twin–inspired AI prescription pipeline. (A) Children undergoing postcycloplegic reexamination were enrolled in a prospective randomized testing sequence, within-subject agreement study (n = 183) and underwent traditional autorefraction, AI wavefront refraction, and masked subjective refraction. Refraction data were converted into power vector components (SE, J0, and J45) for paired comparison, Pearson correlation, Bland–Altman analysis, and absolute SE error assessment. (B) The digital twin–inspired AI prescription pipeline included five steps: wavefront acquisition with image quality assessment (IQA)-gated frame selection, patient-specific virtual optical eye model construction, virtual lens simulation, robust optimization under physiological and fitting perturbations, and neural-adaptation mapping to generate an AI-generated 0.25-D spectacle prescription estimate.

Figure 1.

Continued.

Framework Construction and Prescription Estimation

The patient-specific eye–lens simulation framework was designed as a clinical decision support component for estimating spectacle prescriptions from objective wavefront data. The framework incorporated wavefront acquisition, image quality assessment, patient-specific virtual optical eye model construction, virtual trial-lens simulation, perturbation-aware robust optimization, and AI-based neural-adaptation mapping.

The prescription estimate was not generated by direct conversion of raw wavefront coefficients into sphere, cylinder, and axis. Instead, the framework followed a simulation-based prescription generation pathway: quality-controlled wavefront data were first used to construct an individualized optical representation of the measured eye, candidate sphero-cylindrical prescriptions were then tested through a virtual trial-lens library, retinal image quality metrics were evaluated across candidate prescriptions and prespecified perturbation conditions, and the resulting simulation-derived features were mapped to a subjective refraction–derived prescription endpoint. This design was intended to approximate part of the clinical trial-lens refinement process in silico, rather than to provide a conventional objective refractive reading alone.

Step 1: Wavefront Acquisition and Quality-Gated Frame Selection

Whole-eye optical signals were acquired using a high-resolution microelectromechanical system Shack–Hartmann wavefront sensor. For each eye, two to five consecutive wavefront frames were acquired. Before patient-specific optical model construction, all frames underwent a predefined image quality assessment procedure. The quality control metrics included the proportion of valid Shack–Hartmann spots, spot detection completeness, signal-to-noise ratio, centroid stability, pupil coverage, and wavefront reconstruction residual. Frames were retained for analysis only when they met the following criteria: valid spot proportion ≥75%, spot detection completeness ≥80%, signal-to-noise ratio ≥20, centroid displacement ≤3 pixels (µm pending calibration), and wavefront reconstruction residual ≤1.2 pixels. Frames affected by blinking, incomplete pupil capture, unstable fixation, or insufficient spot detection were excluded before model construction.

Step 2: Patient-Specific Virtual Optical Eye Model Construction

Wavefront surface features were reconstructed using Zernike polynomial analysis. Zernike reconstruction was performed up to the order over a standardized pupil diameter of 3.5 mm. The reconstructed wavefront information, including lower-order refractive components and higher-order aberrations, was used to construct a patient-specific virtual optical eye model. This model was not intended to reproduce the full physiology of the eye; rather, it served as an individualized optical representation for evaluating how candidate sphero-cylindrical lenses would modify retinal image quality in the measured eye.

Step 3: Virtual Trial-Lens Simulation across Sphero-cylindrical Candidates

A virtual trial-lens library containing clinically available sphero-cylindrical combinations was applied to the patient-specific virtual optical eye model. This step was designed to digitally reproduce part of the physical trial-lens process used in subjective refraction. Candidate prescriptions were evaluated by simulating their optical effects on the individualized wavefront model using ray-tracing and optical quality calculation details. The virtual trial-lens library included clinically relevant sphero-cylindrical combinations, with spherical powers ranging from +15.00 to −20.00 D in 0.05-D increments, cylinder powers ranging from 0.00 to −10.00 D in 0.05-D increments in minus-cylinder notation, and cylinder axes sampled from 1° to 180° in 1° increments. For each candidate prescription, retinal image quality metrics were calculated, including point spread function (PSF), modulation transfer function (MTF), and visual Strehl ratio. This simulation step is a central distinction from conventional objective refraction, which primarily measures optical status but does not generally test candidate lens interventions within a patient-specific virtual model.

Step 4: Perturbation-Aware Robust Prescription Optimization

To account for physiologic and fitting-related variability during spectacle wear, the optimization procedure incorporated prespecified perturbation ranges. Pupil diameter was varied around the measured pupil diameter by ±0.5 mm, with the simulated pupil diameter constrained within a clinically plausible range of 3.0 to 6.0 mm. Lens decentration was simulated in both the horizontal and vertical directions within ±1.0 mm from the nominal optical center. When vertex distance variation was included, the vertex distance was varied within ±2.0 mm from the nominal spectacle-fitting value. These perturbation ranges were intended to represent small physiologic and spectacle-fitting variations that may occur during routine pediatric spectacle wear, rather than extreme misalignment conditions.

For each candidate sphero-cylindrical prescription, retinal image quality metrics were recalculated across the prespecified perturbation grid. Rather than selecting only the candidate prescription with the highest simulated optical quality under a single ideal condition, the algorithm used a maximin robust optimization strategy to identify a prescription region with relatively preserved optical performance across the perturbation space. The proprietary weighting parameters and implementation details of the optimization algorithm were not disclosed, but the clinically relevant perturbation ranges used for robustness testing are reported here to improve methodological transparency.

Step 5: Deep Convolutional Neural Network–Based Neural-Adaptation Mapping and Prescription Estimation

The final prescription estimation step incorporated a deep convolutional neural network (DCNN) to map objective optical features to subjective refraction–derived prescription endpoints. The DCNN transformed wavefront-derived features, optical quality metrics, and simulation-derived candidate prescription information into a 0.25-D spectacle prescription estimate, including sphere, cylinder, and axis. The AI output was intended to support clinical decision-making rather than replace subjective clinical judgment.

AI Model Development Dataset and Internal Validation

The DCNN was developed using an independent real-world development dataset of 5630 eyes with paired wavefront measurements and final clinically accepted spectacle prescriptions. This dataset was constructed from routine clinical and dispensing records obtained before the present prospective validation study. The age range of the development dataset was 6 to 70 years. Approximately 63% of eyes were from participants younger than 18 years, and approximately 37% were from adults. Thus, the development dataset was not predominantly adult-derived and included a substantial pediatric component. The refractive distribution of the development dataset covered an SE range of −8.00 to +8.00 D, with cylinder power within 5.00 D. Development and validation splitting was performed at the eye level before model training, and no eye in the internal validation subset was used for model fitting or hyperparameter tuning.

All development cases were acquired using the same AI wavefront refraction platform and wavefront acquisition protocol as used in the present study. For each case, the input data included quality-controlled wavefront measurements, wavefront-derived optical features, optical quality metrics, and simulation-derived candidate prescription information. The output label was defined as the final clinically accepted spectacle prescription. In cases in which the initial AI-generated prescription was accepted after dispensing and wearer feedback, that prescription was retained as the accepted label. In cases with wearer nonacceptance or clinically identified prescription intolerance, an experienced optometrist performed additional review and prescription refinement; the revised prescription was used as the accepted label only if it was subsequently accepted by the wearer. This human-in-the-loop feedback process was used to improve the mapping between objective wavefront-derived features and clinically acceptable subjective prescription endpoints.

The development dataset was divided into training and internal validation subsets at the eye level using a 7:3 split ratio. The present prospective validation cohort of 183 right eyes was not included in model training, hyperparameter tuning, internal validation, or model updating. The model version evaluated in the present study was locked before analysis of the 183-eye validation cohort.

Before prospective validation, the model was evaluated in the internal validation subset using predefined clinical acceptance criteria based on agreement with the final accepted subjective refraction–derived prescription. Model iteration was stopped and the final version was locked when the internal validation results met the prespecified threshold achievement criteria: approximately ≥50% of eyes within ≤0.25 D and ≥90% of eyes within ≤0.50 D for absolute SE error and mean absolute errors for J0 and J45 within approximately 0.05 D. These threshold-based criteria were used as pragmatic internal performance targets before testing the locked model in the independent 183-eye prospective validation cohort.

Model Credibility, Validation, and Uncertainty Considerations

The present study was designed as a prospective clinical validation of the framework against masked subjective refraction rather than as a complete regulatory-grade verification, validation, and uncertainty quantification assessment. Several credibility-oriented elements were incorporated: input quality control through the image quality assessment module, patient-specific optical model construction, prespecified perturbation assumptions for robust optimization, use of an independent AI-development dataset, and prospective agreement testing in an independent pediatric postcycloplegic cohort. Uncertainty in the final prescription estimates was characterized clinically using paired differences, Bland–Altman limits of agreement, and absolute error thresholds. Formal code verification, external multicenter validation, uncertainty propagation from measurement noise and fitting perturbations, and longitudinal feedback-based model updating remain necessary in future studies.

Statistical Analysis

Statistical analysis was performed using SPSS software (Version 23.0; IBM Corp., Armonk, NY, USA). Before statistical analysis, all refractive data were converted into power vector components: SE, J0, and J45. The following formulas were used: SE = sphere + cylinder / 2; J0 = − (cylinder / 2) × cos(2 × axis); and J45 = −(cylinder/2) × sin(2 × axis), with the axis expressed in radians. All prescriptions were analyzed in minus-cylinder notation.

Agreement between each objective method and subjective refraction was assessed for SE, J0, and J45. Paired differences were calculated for traditional autorefraction versus subjective refraction and for AI wavefront refraction versus subjective refraction. Depending on data distribution, paired t-tests or Wilcoxon signed-rank tests were used to compare paired differences. Pearson correlation coefficients were calculated to assess the linear association between each objective method and subjective refraction.

Normality of paired differences was assessed using the Shapiro–Wilk test. When paired differences were not normally distributed, Wilcoxon signed-rank tests were used. Because correlation does not measure agreement, correlation analyses were interpreted as supplementary to Bland–Altman analyses.

Bland–Altman analysis was performed to evaluate agreement between methods, with calculation of mean bias and 95% LoA. To further describe clinically relevant prescription accuracy, absolute SE error was calculated for each objective method relative to subjective refraction. For each absolute SE error threshold (≤0.25 D and ≤0.50 D), each eye was classified as either within or outside the threshold for each objective method relative to subjective refraction. Because traditional autorefraction and AI wavefront refraction were evaluated in the same eyes, paired binary threshold achievement rates were compared using exact McNemar tests. Discordant paired counts were reported to show the direction of method-level differences.

Absolute SE errors between methods were additionally compared using a paired test. To address multiplicity across SE, J0, and J45 comparisons, adjusted P values were calculated using the Holm–Bonferroni method as a sensitivity analysis. Exploratory stratified analyses were performed according to reference-standard subjective SE in 1.00-D intervals: >−1.00 D, −2.00 to −1.00 D, −3.00 to −2.00 D, −4.00 to −3.00 D, −5.00 to −4.00 D, and ≤−5.00 D. Within each stratum, SE bias, 95% LoA, and absolute SE error threshold achievement rates within ≤0.25 D and ≤0.50 D were summarized descriptively for AI wavefront refraction and traditional autorefraction. Because some strata contained relatively few eyes, these analyses were considered exploratory and were not used for confirmatory hypothesis testing. To address multiplicity, Holm–Bonferroni adjustment was applied as a sensitivity analysis across the six paired hypothesis tests involving SE, J0, and J45 for the two objective methods relative to subjective refraction. Both unadjusted and adjusted P values were reported, whereas Bland–Altman bias and limits of agreement were interpreted as the primary agreement metrics.

All statistical tests were two-sided, and a P value <0.05 was considered statistically significant.

Results

Demographic Characteristics

After completion of data screening, 183 right eyes from 183 pediatric participants were included in the final statistical analysis. The mean age of the study cohort was 9.9 ± 2.0 years. Overall, 104 participants were male (56.8%) and 79 were female (43.2%). Autorefraction, AI wavefront refraction, and subjective refraction data from all included right eyes were successfully converted into SE, with-the-rule/against-the-rule astigmatic vector component (J0), and oblique astigmatic vector component (J45). Participant characteristics and data availability are summarized in Table 1.

Table 1.

Participant Characteristics and Data Availability

Characteristic	Overall Cohort
No. of children enrolled/analyzed	183
Age, y	9.9 ± 2.0
Age range, y	6.2–15.7
Male sex	104 (56.8)
Female sex	79 (43.2)
Subjective SE distribution, D	−2.46 ± 1.49
Subjective J0 distribution, D	−0.28 ± 0.32
Subjective J45 distribution, D	0.03 ± 0.13

Values are presented as mean ± standard deviation or number (%) unless otherwise specified.

Wavefront Image Quality Control Yield

Across the 183 local acquisitions, a total of 505 wavefront frames were acquired, of which 303 frames (60.1%) met the predefined quality control criteria and were used for optical modeling. The image quality assessment was performed before AI prescription estimation and without access to subjective refraction results.

Original Refraction Values across Methods

The mean subjective SE was −2.46 ± 1.49 D, compared with −2.26 ± 1.58 D for traditional autorefraction and −2.42 ± 1.53 D for AI wavefront refraction. The mean subjective J0 was −0.28 ± 0.32 D, compared with −0.32 ± 0.34 D for traditional autorefraction and −0.28 ± 0.37 D for AI wavefront refraction. The mean subjective J45 was 0.03 ± 0.13 D, compared with 0.05 ± 0.14 D for traditional autorefraction and 0.03 ± 0.17 D for AI wavefront refraction. Original refraction values and agreement metrics are summarized in Table 2.

Table 2.

Original Values and Agreement Metrics across Refraction Methods

Metric		Subjective Refraction, Mean ± SD, D		Traditional Autorefraction, Mean ± SD, D		AI Wavefront Refraction, Mean ± SD, D
Original Refraction Values by Method
SE		−2.46 ± 1.49		−2.26 ± 1.58		−2.42 ± 1.53
J0		−0.28 ± 0.32		−0.32 ± 0.34		−0.28 ± 0.37
J45		0.03 ± 0.13		0.05 ± 0.14		0.03 ± 0.17
Metric	Method Comparison	Bias ± SD, D	95% CI of Bias, D	P Value	Holm-Adjusted P Value	Pearson r	95% LoA, D
Agreement and Clinical Accuracy Relative to Subjective Refraction
SE	Traditional autorefraction vs. subjective refraction	0.21 ± 0.52	0.13 to 0.28	<0.001	<0.006	0.945	−0.81 to 1.22
SE	AI wavefront refraction vs. subjective refraction	0.04 ± 0.35	−0.01 to 0.09	0.153	0.459	0.972	−0.67 to 0.74
J0	Traditional autorefraction vs. subjective refraction	−0.04 ± 0.21	−0.07 to −0.01	0.006	0.030	0.806	−0.45 to 0.37
J0	AI wavefront refraction vs. subjective refraction	−0.01 ± 0.19	−0.03 to 0.02	0.681	1.000	0.852	−0.39 to 0.37
J45	Traditional autorefraction vs. subjective refraction	0.02 ± 0.13	0.00 to 0.04	0.037	0.148	0.548	−0.23 to 0.27
J45	AI wavefront refraction vs. subjective refraction	0.00 ± 0.15	−0.02 to 0.03	0.725	1.000	0.519	−0.29 to 0.30

Agreement and Error Analysis of Spherical Equivalent

For SE, both traditional autorefraction and AI wavefront refraction showed strong linear correlations with subjective refraction. The Pearson correlation coefficient was 0.945 for traditional autorefraction and 0.972 for AI wavefront refraction.

Compared with subjective refraction, traditional autorefraction showed a mean paired difference of 0.21 ± 0.52 D (95% confidence interval [CI], 0.13–0.28; P < 0.001). The corresponding mean paired difference for AI wavefront refraction was 0.04 ± 0.35 D (95% CI, −0.01 to 0.09; P = 0.153).

Bland–Altman analysis showed a mean SE bias of 0.21 D for traditional autorefraction, with 95% limits of agreement ranging from −0.81 to 1.22 D. For AI wavefront refraction, the mean SE bias was 0.04 D, with 95% limits of agreement ranging from −0.67 to 0.74 D.

For absolute SE error relative to subjective refraction, the proportion of eyes within ≤0.25 D was 100 of 183 eyes (54.6%) for traditional autorefraction and 121 of 183 eyes (66.1%) for AI wavefront refraction. In the paired threshold achievement analysis, 43 eyes were within the ≤0.25-D threshold only with AI wavefront refraction, whereas 22 eyes were within the threshold only with traditional autorefraction. The difference was statistically significant by exact McNemar test (P = 0.013).

For the ≤0.50-D threshold, 159 of 183 eyes (86.9%) were within threshold for traditional autorefraction, and 167 of 183 eyes (91.3%) were within threshold for AI wavefront refraction. In the paired analysis, 20 eyes were within the ≤0.50-D threshold only with AI wavefront refraction, whereas 12 eyes were within the threshold only with traditional autorefraction. This difference did not reach statistical significance by exact McNemar test (P = 0.215).

In exploratory 1.00-D subjective SE stratified analyses, AI wavefront refraction maintained mean SE biases close to zero across all refractive strata, ranging from −0.01 to 0.08 D. The corresponding traditional autorefraction biases ranged from −0.05 to 0.56 D. AI wavefront refraction showed ≤0.50-D absolute SE error rates of 75.0% to 100.0% across strata, whereas traditional autorefraction showed corresponding rates of 81.2% to 92.1% in most strata but lower ≤0.25-D rates in several intermediate myopic ranges. The most myopic stratum (subjective SE ≤−5.00 D) included only 12 eyes and showed wider limits of agreement for both objective methods; this subgroup finding should therefore be interpreted cautiously. Detailed stratified results are provided in Table 3.

Table 3.

Exploratory Stratified SE Agreement Analysis According to 1.00-D Subjective Spherical Equivalent Intervals

Subjective SE stratum	n	Method	Bias, Mean ± SD, D	95% LoA, D	≤0.25 D, n (%)	≤0.50 D, n (%)
≤−5.00 D	12	AI wavefront refraction	0.04 ± 0.37	−0.68 to 0.75	7/12 (58.3)	9/12 (75.0)
		Traditional autorefraction	−0.05 ± 0.44	−0.92 to 0.82	8/12 (66.7)	10/12 (83.3)
−5.00 to −4.00 D	16	AI wavefront refraction	0.08 ± 0.21	−0.33 to 0.48	12/16 (75.0)	16/16 (100.0)
		Traditional autorefraction	0.56 ± 0.98	−1.36 to 2.47	5/16 (31.2)	13/16 (81.2)
−4.00 to −3.00 D	32	AI wavefront refraction	−0.01 ± 0.52	−1.04 to 1.01	20/32 (62.5)	26/32 (81.2)
		Traditional autorefraction	0.29 ± 0.57	−0.84 to 1.41	14/32 (43.8)	27/32 (84.4)
−3.00 to −2.00 D	38	AI wavefront refraction	0.07 ± 0.28	−0.48 to 0.62	21/38 (55.3)	36/38 (94.7)
		Traditional autorefraction	0.07 ± 0.42	−0.76 to 0.90	23/38 (60.5)	35/38 (92.1)
−2.00 to −1.00 D	60	AI wavefront refraction	0.02 ± 0.39	−0.75 to 0.78	42/60 (70.0)	55/60 (91.7)
		Traditional autorefraction	0.19 ± 0.36	−0.52 to 0.89	37/60 (61.7)	52/60 (86.7)
>−1.00 D	25	AI wavefront refraction	0.08 ± 0.20	−0.31 to 0.47	18/25 (72.0)	25/25 (100.0)
		Traditional autorefraction	0.25 ± 0.39	−0.52 to 1.02	13/25 (52.0)	22/25 (88.0)

Agreement and Error Analysis of Astigmatic Vector Components

For J0, both objective methods showed positive correlations with subjective refraction. The Pearson correlation coefficient was 0.806 for traditional autorefraction and 0.852 for AI wavefront refraction.

Compared with subjective refraction, traditional autorefraction showed a mean J0 difference of −0.04 ± 0.21 D (95% CI, −0.07 to −0.01; P = 0.006). The mean J0 difference for AI wavefront refraction was −0.01 ± 0.19 D (95% CI, −0.03 to 0.02; P = 0.681).

In Bland–Altman analysis, the 95% limits of agreement for J0 ranged from −0.45 to 0.37 D for traditional autorefraction and from −0.39 to 0.37 D for AI wavefront refraction.

For J45, the Pearson correlation coefficient with subjective refraction was 0.548 for traditional autorefraction and 0.519 for AI wavefront refraction.

Compared with subjective refraction, traditional autorefraction showed a mean J45 difference of 0.02 ± 0.13 D (95% CI, 0.00 to 0.04; P = 0.037). The mean J45 difference for AI wavefront refraction was 0.00 ± 0.15 D (95% CI, −0.02 to 0.03; P = 0.725).

In Bland–Altman analysis, the 95% limits of agreement for J45 ranged from −0.23 to 0.27 D for traditional autorefraction and from −0.29 to 0.30 D for AI wavefront refraction. The main bias estimates for SE, J0, and J45 are shown in Figure 2, and the corresponding Bland–Altman agreement plots are shown in Figure 3.

Figure 2.

Main agreement results for spherical equivalent and astigmatic vector components. (A–C) Difference plots showing the mean bias and 95% confidence interval for traditional autorefraction and AI wavefront refraction relative to masked subjective refraction for SE (A), J0 (B), and J45 (C). Individual points represent eye-level paired differences. The horizontal dashed line indicates zero bias. (D) Forest plot summarizing the mean bias and 95% confidence intervals across SE, J0, and J45. Compared with traditional autorefraction, AI wavefront refraction showed mean biases closer to zero and confidence intervals closer to zero across refractive components.

Figure 3.

Bland–Altman agreement plots comparing refraction methods. Traditional autorefraction versus subjective refraction for SE (A), J0 (B), and J45 (C). AI wavefront refraction versus subjective refraction for SE (D), J0 (E), and J45 (F). In each panel, the solid horizontal line indicates the mean bias, and the dotted horizontal lines indicate the 95% LoA. Mean bias and 95% LoA values are annotated directly within each panel. For each outcome (SE, J0, and J45), the paired panels were plotted using identical y-axis limits to facilitate visual comparison between methods.

Figure 2.

Continued.

Overall, AI wavefront refraction showed smaller mean differences from subjective refraction across SE, J0, and J45. The most consistent improvement was observed for SE, whereas the agreement pattern for J45 was less definitive.

After Holm–Bonferroni adjustment across the six paired comparisons, the differences for traditional autorefraction remained statistically significant for SE (adjusted P < 0.006) and J0 (adjusted P = 0.030), whereas the J45 difference for traditional autorefraction was no longer statistically significant (adjusted P = 0.148). The paired differences between AI wavefront refraction and subjective refraction remained nonsignificant for SE, J0, and J45 after adjustment.

Discussion

In this prospective, randomized testing sequence, within-subject agreement study, the digital twin–inspired AI wavefront refraction framework showed closer agreement with reference-standard subjective refraction than traditional autorefraction for several clinically relevant refractive parameters. This study should therefore be interpreted as a prospective clinical validation of a patient-specific eye–lens simulation framework, rather than merely as a comparison of two objective refraction devices. Because the AI output was generated after virtual prescription testing and robust optimization within an individualized optical model, the agreement results provide mechanism-informed clinical support for this simulation-based prescription generation pathway.

The most consistent difference was observed for SE, where AI wavefront refraction showed a smaller mean bias, narrower Bland–Altman limits of agreement, and higher absolute SE error hit rates at both the ≤0.25-D and ≤0.50-D thresholds. For J0, AI wavefront refraction also showed a smaller mean difference and slightly narrower limits of agreement. For J45, however, interpretation should be more cautious: although the mean difference was closer to zero for AI wavefront refraction, the Pearson correlation was slightly lower and the limits of agreement were wider than those of traditional autorefraction. This cautious interpretation is further supported by the multiplicity-adjusted analysis, in which the traditional autorefraction J45 difference was no longer statistically significant after Holm–Bonferroni correction. Taken together, these findings suggest that the proposed framework may improve the conversion of objective wavefront information into prescription estimates, particularly for SE, while its performance for oblique astigmatic components requires further evaluation.^17,18

The positive SE bias observed with traditional autorefraction is consistent with previous reports describing discrepancies between objective and subjective refraction in children, particularly under conditions involving accommodation, cycloplegia, or postcycloplegic recovery.^19,20 Conventional objective refraction systems rely on simplified optical assumptions and may be affected by physiologic noise, including accommodative microfluctuation, pupil size variation, tear film instability, and fixation variability.^9,21,22 These factors may be more prominent in pediatric populations because children often have less stable fixation and more limited tolerance for repeated measurements. In this context, the present results support the view that additional computational processing of wavefront data may reduce some, but not all, of the discrepancy between objective measurement and subjective refraction.

Subjective refraction remains the clinical reference standard for spectacle prescription because it integrates optical correction with patient perception and examiner judgment. However, pediatric subjective refraction can be affected by cooperation, attention span, fatigue, and variable responses during lens refinement. The present system should therefore be interpreted as a decision support approach rather than a replacement for subjective refraction or professional judgment. Its potential value lies in providing an objective prescription estimate that is closer to the subjective endpoint, especially for SE, and that may serve as a more informative starting point for clinicians. Whether this translates into shorter examination time, reduced examiner workload, or improved clinical workflow was not directly tested in this study and should be evaluated in future work.

An important aspect of this study is the positioning of the system as a digital twin–inspired computational component rather than a complete ophthalmic digital twin. In medical informatics, a full digital twin generally implies a dynamic virtual counterpart with continuous bidirectional feedback from the biological system.⁵ The framework evaluated here does not yet include real-time physiologic feedback or longitudinal adaptive updating. We have therefore avoided referring to the current system as a complete digital twin and have used the term digital twin–inspired only to indicate patient-specific virtual representation, in silico intervention simulation, and decision support use. Instead, it uses quality-controlled wavefront data to construct a patient-specific optical model and to estimate prescription parameters through optical simulation and optimization. This more cautious definition is important because it clarifies the translational boundary of the method while retaining its relevance to precision refractive computation.^12,23

The optical rationale for the framework also differs from a simple strategy of complete aberration cancellation. Previous work has shown that wavefront-derived image quality metrics, including PSF-, MTF-, and visual Strehl-based approaches, can help predict visual performance and guide optical correction.^24–26 However, a prescription that maximizes retinal image quality under a single static condition may not be optimal under everyday spectacle wear, where small changes in pupil size, lens centration, vertex distance, and gaze direction are common. Misalignment and decentration have been shown to affect the performance of customized or high-order optical corrections.^27–29 The use of robust optimization in the present framework was intended to reduce sensitivity to small perturbations by favoring prescription estimates with relatively preserved simulated image quality across a range of assumed conditions.^30,31 This mechanism provides a plausible explanation for the improved SE agreement, although the present study did not independently test each perturbation parameter and therefore cannot isolate the contribution of any single algorithmic step.

The pediatric postcycloplegic reexamination setting is clinically relevant because cycloplegia is commonly used to reveal refractive status in children, whereas final spectacle prescribing often occurs after partial or complete recovery of physiologic accommodation.^11,19 This creates a practical gap between cycloplegic objective measurement and the subjective endpoint used for dispensing. The proposed framework attempts to address this gap by combining wavefront-derived optical features with neural-adaptation mapping based on matched wavefront and subjective refraction data. The observed ≤0.50-D SE hit rate of 91.3% suggests that the AI-generated prescription estimates were frequently close to subjective refraction in this cohort.^15,24,32 Nevertheless, this result should be interpreted as agreement with an immediate subjective endpoint rather than evidence of long-term wearing comfort or clinical superiority.

From a clinical perspective, the findings may be relevant to pediatric refractive care and myopia management, where accurate baseline refraction is important for spectacle correction, monitoring progression, and selecting optical interventions such as defocus spectacle lenses or orthokeratology.^33,34 The present data suggest that an AI-assisted wavefront framework could provide clinicians with an objective estimate that is more closely aligned with subjective refraction than traditional autorefraction, particularly for SE. Such a tool may support standardization in high-volume clinical environments, but its effect on workflow efficiency, prescribing decisions, patient satisfaction, and downstream treatment outcomes remains to be demonstrated. Clinical decision support systems in ophthalmology require careful validation not only for agreement metrics but also for usability, safety, generalizability, and integration into existing clinical workflows.^35,36

Several limitations should be considered when interpreting these findings. First, this was a single-center study using one experienced optometrist as the reference examiner. This approach reduced interexaminer variability but may limit generalizability to other clinical settings and examiner experience levels. Second, only the right eye of each participant was included in the analysis to avoid intereye correlation; therefore, the results do not address binocular prescription balancing or interocular clinical decision-making. Third, the study evaluated agreement with immediate subjective refraction and did not assess long-term wearing comfort, visual fatigue, adaptation, patient-reported outcomes, or longitudinal myopia-related outcomes. Fourth, the algorithm was evaluated in a selected pediatric population, and its applicability to patients with irregular corneal astigmatism, severe ocular surface disease, media opacity, complex strabismus, amblyopia, or prior orthokeratology use remains uncertain. Fifth, the frame-level quality control acceptance rate was 60.1% (303/505 frames). Although sufficient quality-controlled frames were available for all analyzed eyes, exclusion of low-quality frames may introduce selection bias if frame rejection is associated with younger age, poorer fixation, higher refractive error, or ocular surface instability. The present study did not perform a formal sensitivity analysis using alternative IQA thresholds, and the relationship between frame rejection and patient-level characteristics was not specifically evaluated. Future technical validation should assess the robustness of model performance to IQA threshold settings and frame selection rules. Finally, although the framework includes several computational components, including image quality gating, optical modeling, robust optimization, and neural-adaptation mapping, this study was not designed to determine the independent contribution of each component. Accordingly, the present findings should be regarded as mechanism-informed validation of the final system output, not as component-level mechanistic proof; future ablation studies should compare framework versions with and without virtual lens simulation, perturbation-aware robust optimization, and neural-adaptation mapping.

Future studies should include multicenter prospective validation with broader age ranges, refractive distributions, ethnic backgrounds, and clinical practice settings. Additional endpoints should extend beyond agreement with subjective refraction to include repeatability, examination time, clinician acceptance, patient-reported comfort, contrast sensitivity, adaptation during real-world wear, and longitudinal stability of prescribed correction. Further technical evaluation should also examine the contribution of individual model components, the effect of pupil and decentration assumptions, and performance in eyes with higher-order aberration profiles, high astigmatism, anisometropia, or other complex refractive features not fully characterized in the present development dataset. As AI-assisted medical software moves toward clinical deployment, model version control, data bias monitoring, external validation, and reporting standards consistent with AI clinical research guidelines will be important for ensuring safe and reproducible use.^37,38

Overall, this study suggests that a digital twin–inspired AI wavefront refraction framework can generate pediatric postcycloplegic prescription estimates that are closer to reference-standard subjective refraction than traditional autorefraction for SE and J0, with more nuanced findings for J45. The framework should be viewed as a clinical decision support tool that may help standardize objective to subjective refraction translation, rather than as a substitute for optometrist judgment or direct evidence of improved long-term clinical outcomes.