# paper.pdf ## Metadata - PDFFormatVersion=1.6 - IsLinearized=true - IsAcroFormPresent=false - IsXFAPresent=false - IsCollectionPresent=false - IsSignaturesPresent=false - CreationDate=D:20260710115456+05'30' - ModDate=D:20260710115527+05'30' - Producer=Acrobat Distiller 10.0.0 (Windows) - pdf:keywords="visual field"; "optical coherence tomography"; "artificial intelligence"; "deep learning"; "perimetry" - dc:format=application/pdf - dc:title=Multicenter OCT-Based Visual Field Representations via Segmentation-Free 3D CNNs: Forecasting, Longitudinal Variability, and Progression Detection - dc:identifier=doi:10.1167/tvst.15.7.18 - dc:description=Translational Vision Science and Technology, 2026. doi:10.1167/tvst.15.7.18 - prism:doi=10.1167/tvst.15.7.18 - prism:url=https://doi.org/10.1167/tvst.15.7.18 - prism:aggregationtype=journal - prism:publicationname=Translational Vision Science and Technology - prism:issn=2164-2591 - dc:creator=Makoto Koyama,Hidenori Takahashi,Satoru Inoda,Chihiro Mayama,Yuta Ueno,Yoshikazu Ito,Tetsuro Oshika,Masaki Tanito - dc:subject=visual field,optical coherence tomography,artificial intelligence,deep learning,perimetry - xmpmm:documentid=uuid:690686b0-83c7-37e0-8071-977832cc32cf - xmpmm:instanceid=uuid:12ed5621-8d9d-3947-8f1b-bec91fb310b4 ## Contents ### Page 1 Artificial IntelligenceMulticenter OCT-Based Visual Field Representations via Segmentation-Free 3D CNNs: Forecasting, Longitudinal Variability, and Progression Detection Makoto Koyama1, Hidenori Takahashi2,3, Satoru Inoda3, Chihiro Mayama4, Yuta Ueno5, Yoshikazu Ito5, Tetsuro Oshika5, and Masaki Tanito61 Minamikoyasu Eye Clinic, Kimitsu, Japan 2 Center for Cyber Medicine Research, Institute of Medicine, University of Tsukuba, Tsukuba, Japan 3 Department of Ophthalmology, Jichi Medical University, Shimotsuke, Japan 4 Department of Ophthalmology, Japan Community Health Care Organization Tokyo Shinjuku Medical Center, Tokyo, Japan 5 Department of Ophthalmology, Faculty of Medicine, University of Tsukuba, Tsukuba, Japan 6 Department of Ophthalmology, Shimane University Faculty of Medicine, Izumo, JapanCorrespondence: Makoto Koyama, Minamikoyasu Eye Clinic, 2-8-30 Minamikoyasu, Kimitsu, Chiba 299-1162, Japan. e-mail: minamikoyasuganka@gmail.comReceived: May 29, 2025Accepted: June 10, 2026Published: July 13, 2026Keywords: visual field; optical coherence tomography; artificial intelligence; deep learning; perimetryCitation: Koyama M, Takahashi H, Inoda S, Mayama C, Ueno Y, Ito Y, Oshika T, Tanito M. Multicenter OCT-Based visual field representations via segmentation-free 3D CNNs: Forecasting, longitudinal variability, and progression detection. Transl Vis Sci Technol. 2026;15(7):18, https://doi.org/10.1167/tvst.15.7.18Purpose: To evaluate segmentation-free three-dimensional convolutional neural networks (3DCNNs) that derive visual field (VF)-shaped representations from macular optical coherence tomography (OCT) volumes (OCT-based estimated VF; OCT-VF) and to assess their performance in forecasting future Humphrey field analyzer (HFA) measure- ments and characterizing longitudinal variability in clinical datasets.Methods: The 3DCNN models were trained on 129,007 heterogeneous real-world paired OCT–HFA datasets from 13,366 patients (24,313 eyes) across five institutions using 10-fold cross-validation. Forecasting accuracy for the last HFA visit was evaluated using HFA-only forecasts and hybrid models that combined HFA-anchored intercept with longitudinal slopes derived from both OCT-VF and HFA (AOS-AVG). Longitudinal variability was assessed as residual deviations from linear trends using jackknife resam- pling.Results: The AOS-AVG yielded the lowest forecasting errors across endpoints and signif- icantly reduced mean absolute error compared with HFA-only forecasts (pointwise thresholds: 24-2, 2.45 vs. 2.68 dB; 10-2, 2.44 vs. 2.65 dB; both P < 0.001). OCT-VF showed significantly lower residual variability than HFA (MD: 24-2, 0.79 vs. 1.13 dB; 10-2, 0.87 vs. 1.07 dB; both P < 0.001). Eyes classified as progression-positive by OCT-VF but negative by HFA exhibited greater perimetric variability and steeper HFA MD decline than true- negative eyes.Conclusions: Hybrid forecasting approaches that incorporate OCT-VF–derived slope information can improve prediction of future visual-field outcomes. The lower variabil- ity of OCT-VF representations may also enable detection of progression-related signals that remain obscured by perimetric noise in routine longitudinal monitoring.Translational Relevance: OCT-VF may serve as a complementary tool for precision disease monitoring alongside current automated perimetry.IntroductionVisual field (VF) testing is crucial for diagnosing and monitoring various ocular conditions, particu- larly glaucoma, a leading cause of irreversible blind- ness worldwide.1–3 The Humphrey field analyzer (HFA; Carl Zeiss Meditec, Jena, Germany) remains the clinical standard for VF assessment, but HFA measurements require sustained attention and repeated responses over several minutes and are known to exhibit substantial test–retest variability, particularly inCopyright 2026 The Authors tvst.arvojournals.org | ISSN: 2164-2591 1 This work is licensed under a Creative Commons Attribution 4.0 International License. ### Page 2 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 2advanced disease.3–5 Repeated VF testing is also time- consuming and burdensome for patients and clinics. These characteristics have motivated the search for complementary biomarkers that can enhance longitu- dinal monitoring. Optical coherence tomography (OCT) has trans- formed ophthalmic imaging by providing high- resolution and reproducible measurements of retinal structures that require only brief fixation for several seconds rather than prolonged response- based testing.6 OCT-based structural parameters have been reported to exhibit lower longitudinal variabil- ity than functional testing, particularly for disease monitoring.7 Conventional structural metrics, such as peripapillary retinal nerve fiber layer and macular ganglion cell complex (GCC) thickness, have demon- strated clinically meaningful associations with VF loss.8–11 However, their ability to fully characterize the complex and spatially heterogeneous nature of functional progression remains an area of active inves- tigation. Recent advances in artificial intelligence (AI) have therefore explored using OCT data—ranging from single-layer thickness maps to complete volumet- ric scans—to generate VF-related outputs, including pointwise sensitivity estimates mapped to standard VF test grids (24-2 or 10-2), as well as global functional indexes, through machine learning and deep learn- ing models.12–27 Conceptually, these approaches can be viewed as data-driven dimensionality reduction techniques that transform high-dimensional OCT structural information into lower-dimensional vector representations aligned with functional measurement domains. The target values used for training are derived from HFA, but once learned, the mapping itself constitutes an OCT-based representation that can be analyzed over time; it does not directly measure visual function. A variety of OCT-based estimated visual field (OCT-VF) models have been proposed, ranging from two-dimensional (2D) approaches based on pre- segmented thickness maps or selected B-scans to three- dimensional (3D) models that use the full volume. Although 2D methods offer computational efficiency, they may be influenced by pre-processing choices. Segmentation-based approaches depend on segmenta- tion quality and device-specific implementations,12–15 and B-scan–based approaches inherently provide only partial sampling of the imaged region.16–22 In contrast, 3D convolutional neural network (3DCNN) models28 can operate directly on complete OCT volumes, lever- aging voxel-level information while preserving the full spatial context of the structural data without requir- ing explicit layer summarization.23–25 Related volumet- ric learning frameworks have also been explored for structural–functional association and glaucoma detec- tion, as well as for predicting future OCT struc- ture from prior volumetric scans.27,29 Our previ- ous research showed that training on comprehensive datasets, including various ocular conditions, yielded better performance in estimating VF from 3D OCT images than glaucoma-specific training.26 Building on this foundation, the present study extends the approach to a large, multicenter cohort and focuses on the longitudinal behavior of OCT-derived VF representations. Despite rapid progress in OCT-VF modeling, most prior studies have emphasized cross-sectional accuracy against HFA (for example, mean absolute error at a single time point) and have provided limited infor- mation about how OCT-VF behaves as a longitu- dinal signal.12–26 In particular, it remains unclear whether the reduced noise expected from an algorith- mic, OCT-based representation translates into lower residual variability over time, earlier detection of signif- icant progression trends, and how its trend-based progression signal compares with that of HFA and standard OCT metrics such as macular GCC thick- ness. Moreover, little is known about whether combin- ing longitudinal OCT-VF and HFA information can improve the forecasting of future VF outcomes in routine clinical follow-up. Therefore, in this study, we trained segmentation- free 3DCNN models that map macular OCT volumes to VF-shaped outputs in a heterogeneous, real-world multicenter dataset comprising 89,711 paired 24-2 tests and 39,296 paired 10-2 tests from more than 13,000 patients and evaluated these OCT-derived VF vectors as structure-derived functional biomarkers for longi- tudinal analysis. Specifically, our aims were to: (1) train and validate 3DCNN models for deriving VF- like representations from macular OCT images across a diverse clinical population; (2) assess whether hybrid approaches combining HFA-anchored intercept with slope information derived from OCT-VF, alone or in combination with HFA-derived slopes, can improve forecasting accuracy for future HFA measurements; (3) compare the longitudinal variability of OCT-VF with that of HFA measurements using residual analy- sis to provide a foundational basis for its predic- tive performance; (4) evaluate progression detection characteristics of OCT-VF relative to HFA as a refer- ence standard; and (5) compare OCT-VF progres- sion signals with those from conventional structural OCT metrics (macular GCC thickness) to determine whether volumetric deep learning analysis provides complementary information beyond single-layer thick- ness measurements. This research explores whether this alternative analysis of volumetric OCT data offers ### Page 3 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 3complementary information for disease monitoring that may enhance clinical decision-making when used alongside conventional perimetry. However, validation against clinical outcomes will be necessary to establish definitive clinical utility.Methods Study Design and ParticipantsThe Institutional Review Boards of Shimane University Hospital (IRB No. KS20230719-3, approved on Aug 10, 2023) and Jichi Medical Univer- sity (IRB No. Rindai24-130, approved on February 4, 2025) approved this retrospective multicenter study. Participating institutions included Shimane University Hospital, Shimane, Japan; Jichi Medical University, Tochigi, Japan; JCHO Tokyo Shinjuku Medical Center, Tokyo, Japan; Omiyahamada Eye Clinic, Saitama, Japan; and Minamikoyasu Eye Clinic, Chiba, Japan. This study adhered to the tenets of the Declaration of Helsinki. Because of the retrospective nature of the study, the requirement for informed consent was waived by the respective IRBs. An opt-out approach was used, and study information was publicly available on each institution’s website and premises, allowing patients to decline participation. We retrospectively collected data from patients who underwent macular OCT imaging or HFA testing between August 5, 2004, and February 12, 2025.Inclusion CriteriaWe retrospectively included all consecutive eyes with either (1) at least one macular OCT scan or (2) at least one VF test using the HFA with a 30-2, 24-2, or 10-2 test pattern using Swedish Interactive Thresh- old Algorithm (SITA) strategy. To reflect real-world clinical practice, no exclusions were made based on ocular diagnosis, and both SITA-Standard and SITA- Fast strategies were used for model training.Overview of Datasets and Study DesignWe trained 3DCNN models using 129,007 paired OCT–HFA datasets from 13,366 patients (24,313 eyes) across five institutions within a patient-wise 10-fold cross-validation framework. Independent from the training process, a longitudinal test dataset and a single-center macular GCC subset were constructed for progression and comparative analyses, respectively.Training Data Acquisition and PreprocessingOCT images were acquired using RS-3000, RS- 3000 Advance, RS-3000 Advance2, or Mirante devices (Nidek, Gamagori, Japan) with a 9 × 9 mm volumet- ric scan protocol centered on the macular, performed under both dilated and undilated conditions. This scan area covers the entire macular region. It extends nasally to include part of the optic nerve head and adjacent peripapillary retinal nerve fiber layer, with the nasal edge typically reaching approximately the center of the optic disc. We applied quality control criteria by excluding OCT scans with signal strength index (SSI) <7 and excluding peripheral test points from 30-2 VF tests (trimmed to match the 24-2 pattern). Reliability thresholds for training HFA data were set at <33% for all three indexes: false-positive rate, false-negative rate, and fixation loss rate. This relatively lenient threshold was selected to maximize the training data volume and enhance model robust- ness against real-world artifacts, consistent with previ- ous studies.12,16,26 We paired OCT scans with HFA tests performed within 90 days of the OCT acqui- sition date, selecting the HFA test with the smallest absolute time difference when multiple tests were avail- able. For eyes with both 24-2 and 10-2 test patterns, we included all available data. For eyes with only one test pattern (e.g., only 24-2 but not 10-2) within the 90-day window, we used mask values to indicate missing data points. We included all eligible paired data from eyes with multiple OCT scans and VF tests in the model development dataset, using the raw pointwise thresh- olds and mean deviation (MD) values without any clipping.Model Architecture and ConfigurationWe adopted a segmentation-free 3DCNN model based on the EfficientNet3D-b0 architecture30 with a 30% dropout layer to reduce overfitting (Fig. 1). This architecture was selected for its optimal balance of model complexity and computational efficiency for 3D volumetric data. The model was trained from scratch using a comprehensive OCT dataset. Input OCT images were standardized to a resolution of 224× 224 × 128, and z-score normalization was applied to both images and VF data. The model outputs estimated VF pointwise thresholds (52 points for 24- 2 and 68 points for 10-2 patterns) and their respec- tive MDs. Following previous reports,13,26 we used horizontal flipping for left eye data and vertical flipping as data augmentation techniques. Model training used the Adam optimizer with a mini-batch size of 4 and a variable learning rate schedule, minimizing mean ### Page 4 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 4Figure 1. Schematic representation of the segmentation-free 3DCNN model. We based the model on the EfficientNet3D-b0 architecture with an additional 30% dropout layer, which directly connects to the output layer and 122 outputs: VF pointwise thresholds (52 points for 24-2 and 68 points for 10-2) and their respective MD values.squared error while accounting for missing data points through masked backpropagation.Model Training With 10-Fold Cross-ValidationWe used a 10-fold cross-validation approach using stratified sampling to ensure balanced distribution across folds. Patients were stratified by institution and further divided into 16 subgroups based on test pattern (24-2 or 10-2), MD value (above or below the mean), and number of measurements per eye (above or below the mean). This stratified, patient-wise split- ting ensured that data from the same patient did not appear in more than one set, thereby preventing data leakage while maintaining representative distribu- tions of disease severity and measurement frequency across the training, validation, and test sets (8:1:1 ratio). In 10-fold cross-validation, the entire dataset was divided into 10 equal parts (“folds”). The model train- ing was repeated 10 times, and in each round, a different fold was set aside as the test set, another fold as the validation set, and the remaining eight folds as the training set. By rotating the role of each fold through all combinations, every sample in the dataset was used exactly once for testing, once for validation, and eight times for training. This process resulted in 10 trained models, each optimized on a different combination of training data while being validated and tested on mutually exclusive patient populations. This approach enables a more robust evaluation of model performance than simply using one stratification of test and training data. Impor- tantly, this cross-validation process was used primar- ily to generate the trained models and evaluate cross- sectional estimation performance. The longitudinal variability analysis (described below) was conducted as a separate post-hoc analysis using these trained models.Longitudinal Test Dataset CreationFollowing model training, we constructed a separate longitudinal dataset to evaluate temporal variabil- ity, forecasting performance, and progression-related characteristics of OCT-VF and HFA measurements. The model training dataset was intentionally designed to maximize generalizability and therefore included examinations performed with both SITA-Standard and SITA-Fast strategies, eyes with limited numbers of visits, and examinations with relatively permissive reliability criteria. In contrast, the longitudinal analyses required more consistent testing conditions and suffi- cient repeated observations to allow reliable estima- tion of temporal trends. In addition, stricter reliabil- ity criteria were applied for evaluation, including a lower threshold for false-positive rates. Therefore a dedicated longitudinal dataset was constructed using stricter criteria for test reliability and follow-up struc- ture, as described below. For the OCT-VF dataset, we generated estimates from all available macular OCT images (SSI ≥7) using our trained 3DCNN models. At the patient level, the longitudinal dataset represents a subset of the training dataset: all patients in the longitudinal dataset were also included in the training dataset. Crucially, to prevent data leakage, we assigned each patient deterministically to the specific 3DCNN model where that patient was part of the test fold during the 10-fold cross-validation process. In other words, every OCT-VF estimate was generated exclusively by a model that had never seen that patient’s data during training. For the HFA dataset, examinations not performed using SITA-Standard were excluded to maintain consistency of testing conditions and reduce potential variability, since SITA-Standard is generally consid- ered to have lower test–retest variability than SITA- Fast.5 We excluded only the first HFA test for each eye to reduce learning effects. To improve the relia- ### Page 5 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 5bility of the HFA reference used in the longitudi- nal analysis, we applied specific test selection crite- ria based on an existing guideline and prior evidence. We excluded tests with false-positive rates ≥15% in accordance with the manufacturer’s recommenda- tions for perimetric reliability.31 However, tests were retained regardless of false-negative or fixation loss rates. Elevated false-negative rates are known to corre- late with defect severity in glaucoma rather than reflect- ing poor patient attention alone.32 Likewise, fixation loss has been reported to be a less accurate indicator of perimetric reliability compared with other indexes.33 This strategy—applying a strict false-positive thresh- old while retaining tests with higher false-negative or fixation loss rates—was intended to reduce spurious variability in the HFA reference without systematically excluding eyes with advanced disease. Notably, this false-positive threshold is stricter than that used during model training (<33%), reflecting our intent to use a relatively more reliable subset of HFA measurements for longitudinal reference purposes. We used the point- wise thresholds and MD values without any clipping for both datasets, except that OCT-VF threshold values below 0 dB were set to 0 dB. To minimize within-subject variability, when multi- ple measurements were taken on the same day, we used the last HFA measurement (as repeated testing typically occurs when the first test is unreliable) and calculated SSI-squared weighted averages for OCT- VF measurements. Specifically, for each test point, the weighted average was computed as: Weighted average= ∑ (OCT − VF × SSI2) / ∑ (SSI2)where the summation is over all OCT scans acquired on the same day. SSI-squared weighting (which gives more weight to higher-quality scans while still incorporat- ing information from all available scans) was selected based on preliminary testing comparing linear SSI weighting and unweighted averaging. To align both datasets temporally, when the first and/or last measure- ment dates differed between OCT-VF and HFA, we iteratively trimmed excess measurements in chrono- logical order from the ends of the longer series until the sequences no longer inverted in time. To ensure identical measurement counts per eye between modal- ities, measurements were paired using a greedy match- ing procedure. For each observation in the modal- ity with fewer examinations, the temporally closest measurement from the other modality within ±90 days was selected without replacement based on the smallest absolute inter-test date difference. Observa- tions without a corresponding measurement within 90 days were discarded. Eyes with fewer than six paired measurements were excluded. To characterize sampling density, the distribution of HFA examination counts per eye was examined.Forecasting the Last HFA: Task Definition and ModelsWe evaluated out-of-sample forecasting of the last HFA test for each eye using longitudinal data avail- able strictly prior to the last-visit timepoint. For both pointwise thresholds (24-2 and 10-2 test patterns) and mean deviation (MD), we fitted simple linear trends to the available history and extrapolated to the last HFA measurement. All models were fit separately for each eye and (for pointwise analyses) each test location. The HFA regression forecast (HFA-RF) used only HFA measurements obtained prior to the last HFA. For each eye and location, we regressed the observed HFA value on time using all preceding HFA visits and predicted the last-visit value by linear extrapolation to the last HFA timestamp. For MD, the same procedure was applied to the eye-level MD series. The OCT-VF regression forecast (OCT-RF) used only OCT-VF values strictly earlier than the last HFA. For each eye and location, we regressed the OCT- VF value on time using all OCT-VF tests with times- tamps preceding the last HFA and extrapolated to the last HFA timestamp. For MD, we fit the same linear model to the eye-level OCT-VF MD history. No HFA measurements entered the OCT-RF fit. The HFA-anchored OCT-VF slope forecast (AOS) combines an HFA-anchored intercept with an OCT- VF–derived slope estimated prior to the last HFA. The intercept anchor was defined as the simple average of all HFA measurements preceding the last HFA, thereby preserving the eye- and location-specific functional level while discarding HFA slope information. The slope was estimated from OCT-VF tests preceding the last HFA, as in OCT-RF. The final forecast was obtained by linear extrapolation to the last HFA times- tamp using this anchored intercept and OCT-VF slope. For MD, the same procedure was applied at the eye level. The averaged-slope HFA-anchored forecast (AOS- AVG) extends the AOS framework by incorporating slope information from both modalities. The inter- cept anchor was defined identically to AOS as the simple average of all HFA measurements preceding the last HFA. The slope was computed as the arithmetic mean of the OCT-VF–derived slope and the HFA- derived slope, each estimated using measurements strictly prior to the last HFA. The final forecast was ### Page 6 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 6Figure 2. Study flowchart illustrating patient selection criteria and dataset construction for model training. Starting with 51,451 patients, exclusion criteria were applied, resulting in 13,366 patients included in the training dataset.obtained by linear extrapolation to the last HFA times- tamp using this anchored intercept and averaged slope. For MD, the same procedure was applied at the eye level. For each eye and test location, let tlast denote the timestamp of the last HFA measurement. The four forecasting models are defined as: HFA − RF : ˆy = ¯yHFA + βHFA(tlast − ¯tHFA)OCT − RF : ˆy = ¯yOCT + βOCT(tlast − ¯tOCT )AOS : ˆy = ¯yHFA + βOCT(tlast − ¯tHFA)AOS − AVG : ˆy = ¯yHFA + βOCT + βHFA2(tlast − ¯tHFA)where ¯yHFA and ¯yOCT denote the mean of all prior HFA and OCT-VF values, respectively; βHFA and βOCTare slopes estimated from prior HFA and OCT-VF measurements; and ¯tHFA and ¯tOCT denote the mean timestamps of prior HFA and OCT-VF measurements, respectively. Forecast accuracy was summarized as mean absolute error (MAE) between forecasts and the observed last-visit HFA values, computed per eye and location for pointwise endpoints and per eye for MD. Confidence intervals and paired model compar- isons versus HFA-RF were obtained by patient-level bootstrap (B = 10,000) with paired resampling of all modalities within patient. For multiplicity, we applied the Holm step-down procedure within each endpoint (24-2 pointwise, 10-2 pointwise, 24-2 MD, 10-2 MD) across the three pre-specified comparisons (OCT-RF, AOS, and AOS-AVG vs HFA-RF) and reported Holm- adjusted P values. To harmonize inference across the four endpoints, statistical significance was declared at adjusted P < 0.0125, reflecting an additional Bonfer- roni correction over the four endpoint families. To evaluate performance across disease severity, forecast ### Page 7 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 7Figure 3. Flowchart illustrating patient and data selection for the longitudinal test dataset. Exclusion criteria and the number of excluded records at each step are shown. The resulting test dataset was used for model evaluation.errors (absolute, signed, and squared) were stratified by observed last-HFA values and visualized using boxplots, with overlaid cubic smoothing curves to illustrate severity-dependent trends.Supporting Analyses and ImplementationTo contextualize the forecasting results, we conducted several supporting analyses. Longitudi- nal variability was quantified as residual deviations from eye-specific linear trends using jackknife resam- pling. Mean deviation (MD) progression detection (slope ≤−0.5 dB/year, P < 0.025) was evaluated and compared across modalities. In addition, a single- center subset with concurrent macular GCC measure- ments was analyzed to compare volumetric 3DCNN- derived representations with conventional structural metrics. Finally, for model evaluation, best-available estimate (BAE) visual fields were generated from longitudinal HFA regression models to provide stable, noise-filtered perimetric references. Detailed statistical protocols and error-decomposition procedures are provided in the Supplementary Methods. All statistical tests were two-sided. We performed all analyses using Python 3.12.9 with scikit-learn (1.6.1), statsmodels (0.14.4), and PyTorch (2.6.0). The study workflow and patient selection process are illustrated in Figure 2 (training dataset), Figure 3 (longitudi- nal test dataset), and Supplementary Figure S1 (GCC analysis dataset).Results Training and Test Dataset CharacteristicsTable 1 summarizes the characteristics of the dataset used to train the 3DCNN models. For ### Page 8 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 8Table 1. Training Dataset Characteristics Patient Characteristics Value Total number of patients 13,366 Number of patients in HFA 24-2 12,995 Number of patients in HFA 10-2 4659 Total number of eyes 24,313 Number of eyes in HFA 24-2 23,462 Number of eyes in HFA 10-2 8085 Total number of sets of paired data 129,007 Number of sets of paired data in HFA 24-2 89,711 Number of sets of paired data in HFA 10-2 39,296 Number of HFA 24-2 tests per eye (mean ± SD) 9.69 ± 7.61 Number of HFA 10-2 tests per eye (mean ± SD) 8.09 ± 5.14 Proportion of SITA-Standard strategy in HFA 24-2 (%) 46.5 Proportion of SITA-Standard strategy in HFA 10-2 (%) 47.2 Number of OCT tests per eye (mean ± SD) 5.31 ± 7.27 Mean deviation of the HFA 24-2 (dB, mean ± SD) −5.93 ± 7.65 Mean deviation of the HFA 10-2 (dB, mean ± SD) −8.03 ± 8.69 Age (years, mean ± SD) 65.0 ± 14.2 SSI (mean ± SD) 8.38 ± 1.03 Focus (D, mean ± SD) −2.60 ± 3.56Table 2. OCT-VF and HFA Longitudinal Test Dataset Characteristics Characteristics OCT-VF 24-2 HFA 24-2 OCT-VF 10-2 HFA 10-2 Number of patients 1000 1000 533 533 Number of eyes 1785 1785 934 934 Age (years) 67.3 ± 14.0 67.3 ± 14.0 66.7 ± 14.2 66.8 ± 14.2 First MD (dB) −9.3 ± 8.3* −8.9 ± 8.7 −10.7 ± 9.0* −10.4 ± 9.2 Mean MD (dB) −10.3 ± 8.5* −9.7 ± 8.9 −11.5 ± 9.3* −11.3 ± 9.5 Last MD (dB) −11.1 ± 8.8* −10.6 ± 9.3 −12.3 ± 9.7* −12.1 ± 10.0 Number of tests 8.91 ± 2.71 8.91 ± 2.71 8.18 ± 1.92 8.18 ± 1.92 Inter-modality interval (days) 8.9 ± 23.3 9.0 ± 23.4 2.7 ± 13.6 2.8 ± 13.7 Follow-up duration (days) 2068 ± 1046 2068 ± 1037 1596 ± 576 1599 ± 574 Differences in mean observation periods between OCT-VF and HFA datasets reflect that the temporal alignment criteria were applied at the study level. All inter-modality intervals were constrained to ≤90 days. Inter-modality intervals were defined as the absolute time difference to the nearest examination of the opposite modality. Because nearest-neighbor relationships are directional, OCT→HFA and HFA→OCT values are not necessarily identical. The values are presented as the means ±standard deviations. *MD values for OCT-VF are estimates derived from the OCT-VF models.the longitudinal analysis comparing OCT-VF and HFA variability, we created corresponding time-series datasets for each measurement type from the same patient population. Table 2 summarizes the charac- teristics of these longitudinal test datasets used for comparative analysis. The HFA dataset had mean observation periods of 2068 ± 1037 days (24-2) and 1599 ± 574 days (10-2), whereas the OCT-VF dataset had periods of 2068 ± 1046 days (24-2) and 1596 ±576 days (10-2). There was no significant difference in observation period for 24-2 (P = 0.369, Wilcoxon signed-rank test), whereas a small but statistically significant difference was observed for 10-2 (P < 0.001). To minimize any potential influence of small resid- ual differences in observation period on the compari- son of residual variability, the observation period was ### Page 9 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 9Table 3. Comparison of Within-Modality and Cross-modality Forecasting Errors Test Pattern Modality* Pointwise MAE (dB) MD MAE (dB) 24-2 HFA → HFA 2.68 ± 1.74 1.42 ± 1.86 OCT-VF → OCT-VF 1.24 ± 1.04 0.94 ± 1.22 OCT-VF → HFA 4.02 ± 2.38 2.65 ± 2.61 HFA → OCT-VF 3.80 ± 2.13 2.55 ± 2.45 10-2 HFA → HFA 2.65 ± 1.60 1.33 ± 1.57 OCT-VF → OCT-VF 1.34 ± 0.99 1.03 ± 1.18 OCT-VF → HFA 4.04 ± 2.36 2.52 ± 2.48 HFA → OCT-VF 3.92 ± 2.29 2.49 ± 2.38 Values are presented as mean ± SD. *OCT-VF → OCT-VF: Last OCT-VF predicted from prior OCT-VF measurements. HFA → HFA: Last HFA predicted from prior HFA measurements. OCT-VF → HFA: Last HFA predicted from prior OCT-VF measurements. HFA → OCT-VF: Last OCT-VF predicted from prior HFA measurements.included as one of the covariates in the generalized estimating equations (GEE)34 analysis.Forecasting the Last MeasurementAs shown in Supplementary Figure S2, although the number of examinations per eye was matched between modalities, the distribution decreased at higher visit counts. To evaluate forecasting performance under a general longitudinal scenario, we compared within- modality and cross-modality prediction errors at the last available measurement for each eye using only data strictly prior to that timepoint (Table 3). Across both test patterns (24-2 and 10-2) and both endpoints (pointwise and MD), prediction errors followed a consistent ordering, with the lowest errors for OCT- VF → OCT-VF forecasts, followed by HFA → HFA, HFA → OCT-VF, and the highest errors for OCT-VF→ HFA. For example, in the 24-2 pointwise analy- sis, MAE was 1.24 ± 1.04 dB for OCT-VF → OCT- VF, 2.68 ± 1.74 dB for HFA → HFA, 3.80 ± 2.13 dB for HFA → OCT-VF, and 4.02 ± 2.38 dB for OCT- VF → HFA. Similar patterns were observed for 10-2 and for MD endpoints. Thus, within-modality forecasts yielded lower errors than cross-modality forecasts. In particular, forecasting HFA measurements from OCT- VF resulted in substantially larger errors than forecast- ing HFA from prior HFA measurements. Conversely, predictions targeting OCT-VF showed lower MAE than those targeting HFA across endpoints. Given the higher errors observed in cross-modality forecasting, we next evaluated whether incorporat- ing OCT-VF–derived slope information within an HFA-anchored framework could improve prediction accuracy beyond HFA-only regression for forecast- ing HFA measurements (Fig. 4; Supplementary Table S1). Across both test patterns and endpoints, the averaged-slope model (AOS-AVG) consistently yielded the lowest MAE, followed by AOS and HFA-RF, whereas OCT-RF showed the highest errors. For point- wise thresholds, AOS-AVG demonstrated significantly lower MAE than HFA-RF in both 24-2 (2.45 vs. 2.68 dB, P < 0.001) and 10-2 (2.44 vs. 2.65 dB, P <0.001). AOS also showed modest but statistically signif- icant improvements over HFA-RF in pointwise analy- sis. For MD, AOS-AVG again showed lower MAE than HFA-RF (24-2: 1.32 vs. 1.42 dB, P < 0.001; 10-2: 1.25 vs. 1.33 dB, P = 0.002). In contrast, AOS yielded slightly higher MD errors than HFA-RF in 24-2 (1.51 vs. 1.42 dB, P = 0.006), whereas no significant differ- ence was observed in 10-2 (P = 0.053). OCT-RF, which relies solely on OCT-VF data, showed substantially higher errors across all endpoints. Analysis of forecast error by disease severity (Supplementary Figs. S3–S5) revealed distinct patterns across methods. For pointwise thresholds, AOS and AOS-AVG generally showed lower absolute and squared errors than both HFA-RF and OCT-RF across disease severity levels, with minor exceptions at the highest and lowest sensitivity ranges. In signed error analysis, HFA-RF, AOS, and AOS-AVG showed largely overlapping trends across severity levels for both pointwise thresholds and MD, indicating minimal systematic bias in the slope-based forecasts. AOS and AOS-AVG closely tracked the error pattern of HFA- RF across most sensitivity ranges, although slightly larger errors were observed at the most advanced disease levels. Under both signed and unsigned error ### Page 10 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 10Figure 4. Forecasting the last HFA using OCT-VF and/or HFA histories. (A) 24-2 pointwise MAE for the last HFA, (B) 10-2 pointwise MAE, (C) 24-2 MD MAE, and (D) 10-2 MD MAE. Methods compared: OCT-RF (linear trend fitted using all OCT-VF measurements obtained prior to the last HFA and extrapolated to the last HFA timepoint); AOS (HFA-anchored intercept defined by the mean of all prior HFA measurements combined with the OCT-VF–derived slope estimated from measurements prior to the last HFA); AOS-AVG (mean of all HFA tests prior to the last HFA used as the intercept anchor, with the progression rate defined as the arithmetic mean of OCT-VF–derived and HFA–derived slopes estimated from observations prior to the last HFA, used for extrapolation); HFA-RF (linear trend fitted using all prior HFA measure- ments extrapolated to the last HFA). The AOS approach aims to combine the stable progression rate (slope) derived from OCT-VF with the eye-specific functional intercept provided by HFA. Boxes show the median and interquartile range; whiskers indicate the non-outlier range. Filled circles denote mean MAE; the dashed horizontal line marks the mean MAE of HFA-RF (reference). Asterisks above boxes indicate paired comparisons versus HFA-RF, with P values from a nonparametric patient-level bootstrap (B = 10,000). Multiplicity was controlled within each endpoint (A–D) using the Holm step-down procedure across three pairwise comparisons, comparing OCT-RF, AOS, and AOS-AVG against HFA-RF. Significance codes (display thresholds): **Holm-adjusted P < 0.001 (strong evidence); *Holm-adjusted P < 0.0125 (Bonferroni- derived threshold used for consistency across analyses); ns not significant at P ≥ 0.0125 after Holm adjustment. Lower values indicate better predictions.definitions, OCT-RF demonstrated larger variability and systematic deviation at extreme severity ranges. Spatial analysis of pointwise MAE (Supplementary Fig. S6) showed that AOS-AVG yielded the lowest forecast errors across test locations in both 24-2 and 10-2 patterns, with values comparable to or lower than AOS and uniformly lower than HFA-RF, whereas OCT-RF exhibited the largest errors. Consistent with this observation, severity-stratified analysis by exami- nation count (Supplementary Fig. S7) showed that OCT-RF generally exhibited larger prediction errors across visit strata. Decomposition of squared errors (Supplementary Figs. S8–S9, Supplementary Table S2) showed that the “Remainder” component (residual component) accounted for the largest share of error (38.4% for 24-2; 40.3% for 10-2) and increased with severity. For the 24-2 pattern, spatial analysis indicated that “Location- fixed” bias was most prominent in the temporal region of the field (corresponding to the optic disc area), whereas “Eye-offset” error was higher in the peripheral field.Residual Variability Analysis Using GEEsUsing jackknife resampling, we constructed regres- sion lines for both OCT-VF and HFA measurements separately, and calculated the mean absolute values of the residuals from these lines. Supplementary Table S3 presents the mean absolute residuals and their standard deviations for OCT-VF and HFA datasets. Statistical analysis using GEE showed significantly lower resid- ### Page 11 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 11ual variability for OCT-VF compared to HFA across all four analyses: pointwise thresholds (1.09 vs. 2.39 dB for 24-2, 1.19 vs. 2.37 dB for 10-2) and MDs (0.79 vs. 1.13 dB for 24-2, 0.87 vs. 1.07 dB for 10-2), with all P< 0.001. This pattern was consistently observed across age groups (Supplementary Fig. S10). Variability strat- ified by disease severity (Supplementary Fig. S11) showed that pointwise thresholds generally exhibited lower variability with OCT-VF across the dynamic range, with convergence observed near the 0 dB floor, whereas MD variability showed lower values for OCT- VF in mild-to-severe cases but converged and reversed in advanced cases for both 24-2 and 10-2 patterns. Heat maps of residual variability (Supplementary Fig. S12) showed that OCT-VF had lower variability across all test points compared to HFA.MD Progression Detection and Rate AnalysisTo estimate the false-positive rate under the null hypothesis of no temporal structure, we performed within-eye permutation testing on the longitudi- nal dataset (Supplementary Table S4). The false- positive rates for detecting MD progression (slope≤−0.5 dB/year, P < 0.025) were 0.00727 for HFA 24-2, 0.00560 for OCT-VF 24-2, 0.00750 for HFA 10-2, and 0.00589 for OCT-VF 10-2, all below the nominal 2.5% threshold. Using HFA as the reference standard (Supplemen- tary Table S5), OCT-VF progression detection showed sensitivities of 0.461 (24-2) and 0.396 (10-2), with speci- ficities of 0.917 (24-2) and 0.901 (10-2). Cohen’s kappa values were 0.393 (24-2) and 0.312 (10-2). Detailed confusion matrices are provided in Supplementary Table S6. To further characterize eyes in which OCT-VF detected progression despite non-significant findings on HFA, we compared false-positive (FP) and true- negative (TN) eyes using HFA as the reference standard. FP eyes were defined as eyes showing OCT- VF–detected progression but not meeting HFA signif- icance criteria (OCT-VF+/HFA–), and TN eyes as OCT-VF– / HFA–, based on identical progression criteria (MD slope ≤–0.5 dB/year and P < 0.025). Across both test patterns, FP eyes exhibited signifi- cantly greater HFA MD residual variability, reflected by larger absolute residuals, than TN eyes (24-2: 1.51 vs. 1.08 dB; 10-2: 1.30 vs. 1.05 dB; both P < 0.001; Supplementary Table S7). Density plots of MD resid- uals for FP and TN eyes are provided in Supplemen- tary Figure S13. FP eyes also demonstrated signifi- cantly more negative HFA MD slopes than TN eyes (24-2: −0.412 vs. −0.147 dB/year; 10-2: −0.437 vs.−0.183 dB/year; both P < 0.001; Supplementary Table S8). Corresponding slope distributions are illustrated in Figure 5. Among eyes identified as progressing by both modalities, no statistically significant differ- ence was observed in the time required to detect MD progression between OCT-VF and HFA for either test pattern (Supplementary Fig. S14; Supplementary Table S9). MD progression slopes showed correlations between OCT-VF and HFA (Supplementary Fig. S15). Passing-Bablok regression35 slopes ranged from 0.663 to 1.031 across test patterns and analysis subsets. When the axes were reversed and ordinary least-squares regression was applied (Supplementary Fig. S16), slopes ranged from 0.567 to 0.851. Bland-Altman analysis (Supplementary Fig. S17) showed minimal systematic bias for all eyes, with mean differences in MD slopes (OCT-VF minus HFA) of 0.01 dB/year for 24-2 and 0.03 dB/year for 10-2. Analysis of slope error versus MD (Supplementary Fig. S18) demonstrated no systematic relationship between severity and slope differences, with Passing-Bablok regression slopes near zero across all severity levels (all 95% confidence intervals [CIs] included zero), indicating that OCT-VF showed no systematic bias across the dynamic range of visual field damage. Comparison of mean MD slopes between modal- ities for all eyes (Supplementary Table S10) showed no statistically significant differences between OCT- VF and HFA. For the 24-2 pattern, mean slopes were −0.316 dB/year (OCT-VF) and −0.325 dB/year (HFA), with a mean difference () of 0.010 dB/year (95% CI: −0.023 to 0.043; P = 0.557). For the 10-2 pattern, mean slopes were −0.359 and −0.393 dB/year, respectively ( = 0.034 dB/year; 95% CI: −0.017 to 0.083; P = 0.185). These findings indicate no system- atic difference in estimated progression rates between modalities.Comparison With Macular GCC ProgressionWe performed a subanalysis on a single-center dataset of eyes where longitudinal macular GCC thick- ness measurements were available to compare the progression detection performance of OCT-VF against standard structural metrics (Supplementary Fig. S1). This dataset included 1078 eyes (24-2) and 402 eyes (10- 2) (Supplementary Table S11). Within-eye permutation testing showed false-positive rates below the nominal 2.5% threshold for progression detection based on both MD and GCC slopes (Supplementary Table S12). Given that our 9 × 9 mm macular OCT scan area extends nasally to approximately the optic disc center and thereby captures the retinal ganglion cell axon pathways corresponding to central VF locations, we ### Page 12 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 12Figure 5. HFA MD slopes for FP and TN eyes. FP and TN eyes were defined using the same progression criteria (MD slope ≤ −0.5 dB/year andP < 0.025) with HFA as the reference standard; OCT-VF progression was evaluated using identical thresholds. FP corresponded to OCT-VF+, HFA–, and TN to OCT-VF–, HFA– (see Supplementary Table S6 for full confusion matrices). Panels A (24-2) and B (10-2) display probability density histograms of HFA MD slopes for FP eyes (n = 125 for 24-2, n = 77 for 10-2) and TN eyes (n = 1380 for 24-2, n = 698 for 10-2). FP eyes showed significantly more negative MD slopes, indicating faster functional deterioration that did not satisfy HFA’s progression threshold despite being detected by OCT-VF. All between-group differences were statistically significant (all P < 0.001; see Supplementary Table S8).prioritized the analysis of the 10-2 test pattern. Using HFA as the reference standard, OCT-VF demonstrated better agreement in progression detection compared to GCC across all threshold definitions (Supplementary Tables S13–S14). For 10-2 with the −0.5 μm/year GCC threshold, which substantially exceeds the reported rate of normal age-related GCC thinning (approxi- mately −0.17 μm/year),36 OCT-VF showed sensitivity of 0.753, specificity of 0.875, and Cohen’s κ of 0.560, whereas GCC showed sensitivity of 0.370, specificity of 0.690, and Cohen’s κ of 0.046. The difference in Cohen’s κ between OCT-VF and GCC was 0.512 (95% CI, 0.372 to 0.647; P < 0.001), indicating significantly better agreement for OCT-VF. Similar patterns were observed for 24-2 and across different GCC thresholds. Venn diagram analysis (Supplementary Figs. S19– S21) revealed that the choice of GCC threshold influ- enced progression detection characteristics. At the<0.0 μm/year threshold, GCC demonstrated elevated false-positive rates, whereas the <−1.0 μm/year thresh- old resulted in reduced sensitivity. For 10-2 at the intermediate −0.5 μm/year threshold, GCC perfor- mance was inferior to OCT-VF for both true positives (27 versus 55 eyes) and false positives (102 versus 41 eyes), demonstrating OCT-VF’s better specificity while maintaining higher sensitivity. A similar pattern was observed in the 24-2 analysis. Finally, we evaluated the longitudinal agreement between the rates of change in HFA MD and those derived from OCT-VF and GCC (Supplementary Table S15). OCT-VF MD slopes showed significantly stronger correlations with HFA MD slopes than GCC thickness slopes. For the 10-2 pattern, Pearson’s r was 0.850 for OCT-VF compared with 0.290 for GCC (r = 0.549; 95% CI, 0.351 to 0.772; P < 0.001). Similar results were observed for the 24-2 pattern and when using Spearman’s rank correlation. Disease severity–stratified analyses further showed that OCT- VF maintained stronger associations with HFA MD slopes than GCC across severity levels (Supplementary Table S16). For the 10-2 pattern, Pearson’s r for OCT- VF ranged from 0.732 to 0.866 across severity strata, whereas GCC correlations ranged from 0.189 to 0.540. Scatterplots with Passing–Bablok regression (Supple- mentary Figs. S22, S23) provided visual confirmation that OCT-VF progression rates aligned more closely with HFA MD slopes, whereas GCC slopes exhibited weaker correspondence with functional decline.Agreement Between BAE and OCT-VFSupplementary Table S17 shows the agreement between OCT-VF and the BAE for each parameter. The mean absolute error (MAE) between OCT-VF and BAE was 3.33 dB for 24-2 pointwise thresholds, 3.47 dB for 10-2 pointwise thresholds, 2.27 dB for 24- 2 MD, and 2.25 dB for 10-2 MD. Figure 6 shows the relationship between OCT-VF and the corresponding BAE values for both pointwise thresholds and MDs. Analysis of the absolute error relative to BAE (Supple- mentary Fig. S24) revealed that absolute errors showed increasing trends in mild to moderate cases, followed by a reduction in the most advanced range, resulting in an inverted-U–shaped pattern. Furthermore, analy- sis of the error (OCT-VF minus BAE) relative to BAE (Supplementary Fig. S25) showed a shrinkage toward the central tendency; OCT-VF tended to overesti- mate more severe (lower) thresholds and underestimate better (higher) thresholds. The Bland-Altman analy- sis showed absolute mean differences between OCT- ### Page 13 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 13Figure 6. Correlations between BAE and OCT-VF parameters. (A) 24-2 pointwise thresholds, (B) 10-2 pointwise thresholds, (C) 24-2 MD, and (D) 10-2 MD. Each panel displays boxplots of OCT-VF estimates for each corresponding BAE value, along with the identity line (x =y). BAE values were derived from linear regression over time for each eye. For pointwise thresholds, a separate regression line was fitted for each test location, and the predicted HFA value at each OCT scan date was used as the BAE threshold. For MD, a regression line was fitted using all longitudinal HFA measurements for each eye, and the predicted MD at each OCT scan date was used as the BAE MD. BAE values represent smoothed estimates derived from the longitudinal HFA series and are used as a reference for comparison rather than as independent measurements. OCT-VF values were generated for all available OCT scans in the longitudinal dataset, and each prediction was paired with the corresponding BAE value obtained at that same OCT date.VF and BAE values of 0.6 dB or less across all analy- ses (Supplementary Fig. S26). The model’s estimation accuracy remained relatively stable across various OCT focus values; regression slopes were close to zero and 95% confidence intervals included zero for most analy- ses, indicating minimal impact of refractive errors on performance (Supplementary Fig. S27). The spatial distribution of MAE indicated lower error rates in nasal VF locations in severe cases (Supplementary Figs. S28, S29). When the same trained model was evaluated in the GCC analysis dataset, the observed MAE values between OCT-VF and BAE were numerically smaller (MAE: 2.05 dB for 24-2 thresholds, 2.79 dB for 10-2 thresholds, 1.39 dB for 24-2 MD, and 1.78 dB for 10-2 MD) than those in the full longitudinal cohort (Supple- mentary Tables S17, S18).Analysis of Visual Field Measurements at the Lower Threshold LimitSupplementary Figure S30 shows that OCT-VF produced fewer measurements clustered at 0 dB compared to BAE, with OCT-VF estimates typically remaining slightly above the threshold floor. To further explore this discrepancy, longitudinal analysis (Supple- mentary Fig. S31) showed that HFA measurements were more likely to remain at 0 dB over time. In contrast, OCT-VF estimates tended to fluctuate slightly above the threshold floor. When recovery from 0 dB did occur, HFA values exhibited large and variable increases (24-2: 10.5 ± 7.1 dB, 10-2: 10.4 ± 7.4 dB), while OCT-VF changes were small and stable (24-2: 1.7± 1.6 dB, 10-2: 2.0 ± 1.8 dB).DiscussionOur forecasting analyses provide several insights into the longitudinal behavior of OCT-derived visual field representations. As expected, direct cross- modality forecasting—such as predicting future HFA values using OCT-VF measurements alone—resulted in substantially higher errors than within-modality ### Page 14 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 14predictions, indicating that OCT-derived representa- tions and functional perimetric measurements retain modality-specific variability and are not directly interchangeable at the individual level. However, integrating both modalities within a hybrid framework significantly improved prediction accuracy. In partic- ular, the averaged-slope model (AOS-AVG), which combines an HFA-anchored intercept with the mean of OCT-VF–derived and HFA-derived longitudinal slopes, consistently yielded the lowest forecasting errors. One possible explanation is that averaging slope estimates derived from two different modalities may mitigate modality-specific noise, thereby producing a more stable and accurate estimate of longitudinal change. These findings suggest that OCT-derived information may provide complementary longitudinal signals beyond those captured by perimetry alone, and that integrating slope estimates from both modalities can further stabilize forecasting of future visual-field outcomes. The performance of the AOS framework provides further insight into the mechanism underlying this hybrid forecasting approach. By combining an HFA- anchored intercept with an OCT-VF–derived slope, this framework separates the roles of functional level anchoring and trend estimation. HFA measurements remain essential for anchoring the individual functional level, whereas OCT-VF provides slope information that reflects structure-derived functional change with lower longitudinal variability. This interpretation is consis- tent with our error decomposition analysis, which identified systematic components such as eye-offset errors and location-fixed errors that can be mitigated through intercept anchoring strategies. Interestingly, although the basic AOS model showed no advantage (and even slightly higher error) for MD forecasting, the AOS-AVG model demonstrated significant improve- ments. This indicates that even for spatially averaged indices such as MD, which already exhibit inherently reduced variance, incorporating slopes derived from OCT-based functional estimates provides a stabilizing effect against perimetric noise. Furthermore, the forecasting experiments revealed that prediction difficulty differed depending on the prediction target (OCT-VF vs. HFA). Predictions targeting OCT-VF generally showed smaller errors than those targeting HFA. This difference likely reflects the larger test–retest variability of standard automated perimetry compared with OCT-VF– derived representations. Because OCT-VF represents a structure-derived functional estimate with less variabil- ity, forecasts may be more stable when OCT-VF is used as the prediction target. In contrast, the larger variabil- ity of perimetric measurements increases uncertainty in model estimation, which may amplify forecasting errors when HFA is used as the prediction target. More fundamentally, direct cross-modality prediction—targeting HFA using only OCT-VF estimates—is inherently challenging because each modality contains modality-specific sources of variability. On the HFA side, measurements are influ- enced by patient-specific response behaviors (response bias), which were identified as "eye-offset" errors (individual intercept bias) in our decomposition analy- sis. Conversely, OCT-VF estimates are algorithmically generated and may be subject to case-specific estima- tion biases or a model-driven tendency to regress toward central tendencies. These findings suggest that although HFA remains the clinical reference standard, its measurement noise represents an inherent limitation for forecasting, one that may be partially mitigated by incorporating the relatively stable signal provided by OCT-VF. Our analysis revealed important considerations when estimating visual field measurements near the lower threshold limit. OCT-VF produced fewer measurements at 0 dB compared to the BAE derived from HFA, and longitudinal analysis showed that OCT-VF estimates tended to fluctuate slightly above the floor with smaller variance than HFA. This tendency to underestimate severe damage likely stems from mean squared error-based training, which inher- ently encourages predictions toward central tendencies rather than extreme floor values. Consistent with this floor behavior, analysis of severity-stratified forecasting errors showed a subtle tendency toward larger signed errors for AOS than HFA-RF at test locations with observed thresholds of 0 dB. This pattern may reflect the tendency of OCT-VF estimates to remain slightly above the absolute floor when the functional sensitivity is already at 0 dB. An additional contributing factor may be the increased measurement variability known to occur at low sensitivity levels in standard automated perimetry.4,37 This increased variability in the lower dynamic range elevates the likelihood of perimetric measurements being recorded at the threshold floor, a phenomenon that algorithmically derived estimates such as OCT-VF may inherently smooth over. The better forecasting performance of the hybrid approach can be attributed, at least in part, to the lower measurement variability of OCT-VF. Jackknife resam- pling and GEE analysis demonstrated significantly lower residual variability for OCT-VF across both pointwise thresholds and MD values, with this advan- tage persisting across regions and age groups. A poten- tial concern with neural network-based representations is that models trained on paired OCT–HFA data may regress predictions toward central tendencies rather ### Page 15 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 15than faithfully tracking longitudinal change. Because OCT-VF predictions were generated independently for each visit without explicit temporal smoothing, such regression effects would be expected to manifest as attenuation of estimated slopes. Our analysis directly addressed this possibility: comparison of mean MD slopes between OCT-VF and HFA showed no statis- tically significant differences across test patterns or analysis subsets, with bootstrap confidence intervals for MD slope consistently spanning zero. Furthermore, no significant relationship was observed between HFA MD and MD slope, as Passing–Bablok confidence intervals consistently included zero. Together, these results provide no evidence that OCT-VF progres- sion estimates systematically attenuate longitudinal change or introduce severity-dependent bias. Although Passing–Bablok regression comparing MD progres- sion slopes between HFA and OCT-VF yielded slopes below unity, this pattern is expected under asymmet- ric measurement variability. Greater dispersion in HFA MD slopes increases variance along the predictor axis, which can mechanically reduce regression slopes without implying biological or algorithmic flattening. Importantly, reversing the axes and applying ordinary least-squares regression yielded similarly sub-unity slopes, confirming that slope magnitude is driven by the variance structure of the two measurements rather than by modality-specific attenuation. In the present study, slope analyses focused on global MD trends. Direct pointwise slope compar- isons were not performed because location-specific regression estimates exhibited substantially greater variability than global indices, limiting the stability and interpretability of slope-based statistical compar- isons. Consequently, although no systematic bias was observed at the MD level, localized differences in pointwise slope behavior cannot be entirely excluded. Nevertheless, the forecasting analyses provide indirect context for interpreting this limitation. Although signed pointwise prediction errors were broadly similar between HFA-RF and AOS across disease sever- ity levels, absolute prediction errors tended to be smaller for AOS across most sensitivity ranges. This behavior is compatible with OCT-VF–derived slopes capturing longitudinal variation relevant to point- wise forecasting without introducing large systematic distortions. In progression detection analyses, eyes classified as OCT-VF positive but HFA negative showed signifi- cantly greater perimetric variability and more negative HFA MD slopes than true-negative eyes. These findings suggest that at least some discordant classi- fications may represent eyes with underlying declin- ing tendencies that remain undetected by conventional perimetric progression criteria due to measurement variability. Increased variability reduces the statistical power of perimetric progression tests, making it more difficult for true declining trends to exceed significance thresholds. Although structural–functional dissocia- tion remains a possible explanation, the observed pattern is also compatible with OCT-VF captur- ing progression-related signals that are obscured by variability in perimetric measurements. Consequently, some cases classified as OCT-VF false positives relative to HFA may instead reflect limitations in the sensitiv- ity of perimetric progression detection. This interpre- tation may also explain why eyes classified as progress- ing by both modalities did not show earlier detection by OCT-VF, as such eyes may represent cases with sufficiently strong perimetric progression signals to be detected by conventional tests. The comparison with conventional structural metrics suggested that OCT-VF may offer advantages for monitoring concurrent disease progression. In our subset analysis, OCT-VF progression rates showed stronger alignment with functional decline (HFA MD slopes) than macular GCC thickness slopes across disease severities, whereas GCC thinning demon- strated only modest concordance with central visual field change. Several methodological features may help explain this difference. First, conventional structural metrics reduce the macular volume to a limited set of thickness summaries across retinal layers, whereas the 3DCNN leverages the full volumetric information from the entire macular cube, potentially capturing spatial patterns that are not expressed in aggregated thickness values. Second, because GCC thickness is susceptible to floor effects, OCT-VF appears to preserve useful structure–function signal further into advanced disease. Third, OCT-VF is trained directly to predict functional patterns from paired datasets rather than relying solely on anatomical correlation. These results suggest that volumetric deep learning extracts progression signals that are less apparent in thickness- based metrics, while complementing—rather than replacing—conventional structural OCT measures in longitudinal monitoring. Direct comparison of MAE or RMSE values across published OCT–VF studies requires caution because both the reference standard and the under- lying case mix differ substantially between cohorts. In our study, we used regression-based BAE rather than single-timepoint HFA measurements as the refer- ence standard, motivated by the increase in perimet- ric variability at low sensitivities.4,37 Regression-based estimates leverage routinely collected longitudinal data to provide more stable reference values across the full dynamic range of disease severity, without requiring ### Page 16 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 16repeat perimetry within short intervals or discarding severely damaged test locations. These methodological choices, together with cohort composition, complicate cross-study interpretation of global MAE. For example, when we applied the same trained model to two different cohorts, the GCC analysis subset—enriched with milder disease—yielded lower MAE than the more heterogeneous longitudi- nal dataset, despite identical architecture and weights. Populations enriched with mild disease naturally produce lower error values, whereas regression-based references reduce apparent error by attenuating short- term noise and smoothing low sensitivities. Thus, lower MAE does not necessarily imply a more accurate model, and higher MAE does not necessar- ily indicate inferior performance. Consistent with this, we observed an inverted-U pattern in error relative to BAE—peaking in moderate stages and decreas- ing again in the most advanced range—contrasting with studies that used single raw HFA values as ground truth.16,19,20,22,23,25 Our model also differs from most prior work in being trained on a heterogeneous, multi-condition population rather than glaucoma- specific cohorts.12–25 When considered together— reference standard, severity distribution, and training population—a simple numerical comparison of global MAE across studies risks attributing differences driven by methodology to model superiority or inferiority. Severity-stratified reporting and standardized evalua- tion frameworks will be essential for meaningful cross- study comparisons going forward. This study has several limitations. First, discor- dance between OCT-VF and HFA may arise from model inaccuracies, inherent perimetric noise, or true structural–functional dissociation. Second, analyses were restricted to a single OCT platform with a 9× 9 mm macular scan protocol; generalizability to other devices or acquisition settings requires further validation. Third, residual learning effects and imper- fect temporal alignment between modalities may have added noise to the longitudinal analyses. Finally, the strictly matched longitudinal conditions used in this study—designed to ensure a rigorous compari- son by equalizing measurement counts and matching follow-up periods—may not fully reflect routine clini- cal practice, in which OCT is often acquired more frequently than perimetry. Future research should address these limitations through external validation, prospective longitudinal studies, and broader functional outcomes. Validation across diverse populations, OCT platforms, and clini- cal settings will be essential to establish generalizabil- ity. Prospective studies with standardized follow-up may clarify the temporal relationship between struc- tural and functional change and determine whether improvements in forecasting behavior translate to clinically meaningful benefits. In addition, evalua- tion of multimodal monitoring strategies and case- specific clinical decision frameworks, together with systematic testing across different OCT hardware, will be important for defining where OCT-VF offers the greatest added value and how it can be safely implemented. Notably, in routine follow-up, structural imaging is typically performed more frequently than perimetry owing to its shorter acquisition time and minimal reliance on patient engagement. Furthermore, whereas perimetric monitoring commonly requires exclusion of initial tests to mitigate learning effects, OCT-derived representations can be utilized from the first visit. These practical differences suggest that the complementary value of OCT-VF—including progres- sion detection characteristics and forecasting stabil- ity under higher sampling density—may differ from that observed under the strictly matched analytical conditions of the present study. Prospective head- to-head comparisons preserving real-world sampling patterns will be necessary to quantify these potential effects.ConclusionsA hybrid forecasting framework combining OCT- VF–derived slope information with an HFA-based functional intercept improved longitudinal visual-field prediction, with the averaged-slope model (AOS- AVG) achieving the lowest forecasting errors across endpoints. OCT-VF–derived slopes provided comple- mentary progression information without introduc- ing systematic bias at the global MD level and exhibited lower longitudinal variability than perimet- ric measurements. In addition, OCT-VF showed the potential to highlight progression-related signals in some eyes where high perimetric variability limited conventional progression detection. These findings suggest that structure-derived functional representa- tions may complement standard perimetry for longi- tudinal monitoring. Further prospective studies are needed to determine how such hybrid approaches influ- ence clinical decision-making and patient outcomes.AcknowledgmentsThe authors acknowledge the use of ChatGPT 5.2 (OpenAI) for English language editing and proofread- ### Page 17 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 17ing on Mar 15, 2026. All content was reviewed and verified for accuracy by the authors. Supported by the Ministry of Education, Culture, Sports, Science and Technology of Japan, Grant-in- Aid for Scientific Research (21K16903). The sponsor or funding organization had no role in the design or conduct of this research.Author Contributions: M.K., Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data Curation, Writing – Original Draft, Visualization, Project administration;H.T. Supervision, Formal analysis, Validation, Inves- tigation, Writing – Review & Editing; S.I., Valida- tion, Writing – Review & Editing; C.M., Investiga- tion, Writing – Review & Editing; Y.U., Validation, Writing – Review & Editing; Y.I., Writing – Review & Editing; T.O., Supervision, Writing – Review & Editing; M.T., Supervision, Investigation, Writing – Review & Editing.Data Availability Statements: The datasets gener- ated and/or analyzed during the current study are not publicly available due to privacy and ethical consider- ations. The study was conducted on an opt-out basis, and explicit consent for data sharing was not obtained from all participants. Disclosure: M. Koyama, DeepEyeVision Inc. (C, F, S); H. Takahashi, DeepEyeVision Inc. (C, F, S);S. Inoda, None; C. Mayama, None; Y. Ueno, None;Y. Ito, None; T. Oshika, None; M. Tanito, NoneReferences1. Tham YC, Li X, Wong TY, Quigley HA, Aung T, Cheng CY. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmol- ogy. 2014;121:2081–2090. 2. Weinreb RN, Aung T, Medeiros FA. The patho- physiology and treatment of glaucoma: a review.JAMA. 2014;311:1901–1911. 3. Delgado MF, Nguyen NTA, Cox TA, et al. Automated perimetry: a report by the Ameri- can Academy of Ophthalmology. Ophthalmology. 2002;109:2362–2374. 4. Heijl A, Lindgren A, Lindgren G. Test-retest vari- ability in glaucomatous visual fields. Am J Ophthal- mol. 1989;108:130–135. 5. Artes PH, Iwase A, Ohno Y, Kitazawa Y, Chauhan BC. Properties of perimetric threshold estimates from Full Threshold, SITA Standard, and SITA Fast strategies. Invest Ophthalmol Vis Sci. 2002;43:2654–2659. 6. Huang D, Swanson EA, Lin CP, et al. Optical coherence tomography. Science. 1991;254(5035): 1178–1181. 7. Wu JH, Moghimi S, Nishida T, et al. Evaluation of the long-term variability of macular OCT/OCTA and visual field parameters. Br J Ophthalmol. 2024;108:211–216. 8. Nouri-Mahdavi K, Fatehi N, Caprioli J. Longitu- dinal macular structure-function relationships in glaucoma and their sources of variability. Am J Ophthalmol. 2019;207:18–36. 9. Mohammadzadeh V, Rabiolo A, Fu Q, et al. Lon- gitudinal macular structure-function relationships in glaucoma. Ophthalmology. 2020;127:888–900. 10. Kim S, Lee JY, Kim SO, Kook MS. Macu- lar structure-function relationship at various spa- tial locations in glaucoma. Br J Ophthalmol. 2015;99:1412–1418. 11. Zhang X, Parrish RK, 2nd, Greenfield DS, et al. Predictive factors for the rate of visual field pro- gression in the Advanced Imaging for Glaucoma Study. Am J Ophthalmol. 2019;202:62–71. 12. Asaoka R, Xu L, Murata H, et al. A joint multitask learning model for cross-sectional and longitudinal predictions of visual field using OCT. Ophthalmol Sci. 2021;1(4):100055. 13. Hashimoto Y, Kiwaki T, Sugiura H, et al. Predict- ing 10-2 visual field from optical coherence tomog- raphy in glaucoma using deep learning corrected with 24-2/30-2 visual field. Transl Vis Sci Technol. 2021;10(13):28. 14. Xu L, Asaoka R, Kiwaki T, Murata H, Fujino Y, Yamanishi K. PAMI: a computational mod- ule for joint estimation and progression predic- tion of glaucoma. In: KDD ’21: Proceedings of the ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining. New York: Asso- ciation for Computing Machinery; 2021:3826– 3834. 15. Tong J, Alonso-Caneiro D, Kalloniatis M, Zangerl B. Prediction of visual field defects from mac- ular optical coherence tomography in glaucoma using cluster analysis. Ophthalmic Physiol Opt. 2022;42:948–964. 16. Kamalipour A, Moghimi S, Khosravi P, et al. Deep learning estimation of 10-2 visual field map based on circumpapillary retinal nerve fiber layer thickness measurements. Am J Ophthalmol. 2023;246:163–173. 17. Kihara Y, Montesano G, Chen A, et al. Policy- driven, multimodal deep learning for predicting ### Page 18 OCT-Based Visual Field Forecasting TVST | July 2026 | Vol. 15 | No. 7 | Article 18 | 18visual fields from the optic disc and OCT imaging.Ophthalmology. 2022;129:781–791. 18. Lazaridis G, Montesano G, Afgeh SS, et al. Pre- dicting visual fields from optical coherence tomog- raphy via an ensemble of deep representation learners. Am J Ophthalmol. 2022;238:52–65. 19. Huang X, Sun J, Majoor J, et al. Estimating the severity of visual field damage from retinal nerve fiber layer thickness measurements with arti- ficial intelligence. Transl Vis Sci Technol. 2021; 10(9):16. 20. Mariottoni EB, Datta S, Dov D, et al. Artificial intelligence mapping of structure to function in glaucoma. Transl Vis Sci Technol. 2020;9(2):19. 21. Hemelings R, Elen B, Barbosa-Breda J, et al. Pointwise visual field estimation from optical coherence tomography in glaucoma using deep learning. Transl Vis Sci Technol. 2022;11(8):22. 22. Wong D, Chua J, Bujor I, et al. Comparison of machine learning approaches for structure- function modeling in glaucoma. Ann N Y Acad Sci. 2022;1515:237–248. 23. Chen Z, Shemuelian E, Wollstein G, Wang Y, Ishikawa H, Schuman JS. Segmentation-free OCT-volume-based deep learning model improves pointwise visual field sensitivity estimation. Transl Vis Sci Technol. 2023;12(6):28. 24. Yu HH, Maetschke SR, Antony BJ, et al. Esti- mating global visual field indices in glaucoma by combining macula and optic disc OCT scans using 3-dimensional convolutional neural networks.Ophthalmol Glaucoma. 2021;4:102–112. 25. Mohammadzadeh V, Vepa A, Li C, et al. Predic- tion of central visual field measures from macular OCT volume scans with deep learning. Transl Vis Sci Technol. 2023;12(11):5. 26. Koyama M, Ueno Y, Ito Y, Oshika T, Tanito M. Automated learning of glaucomatous visual fields from OCT images using a comprehensive, segmentation-free 3D convolutional neural net- work model. Sci Rep. 2025;15:13395. 27. George Y, Antony BJ, Ishikawa H, Wollstein G, Schuman JS, Garnavi R. Attention-guided 3D-CNN framework for glaucoma detection and structural-functional association using vol- umetric images. IEEE J Biomed Health Inform. 2020;24:3421–3430. 28. Kopuklu O, Kose N, Gunduz A, Rigoll G. Resource Efficient 3D Convolutional Neural Net- works. In: 2019 IEEE/CVF International Confer- ence on Computer Vision Workshop (ICCVW). Pis- cataway, NJ: IEEE; 2019. 29. Hassan ON, Sahin S, Mohammadzadeh V, et al. Conditional GAN for prediction of glaucoma pro- gression with macular optical coherence tomog- raphy. In: Advances in Visual Computing. Berlin: Springer International Publishing; 2020:761–772. 30. Shi J. EfficientNet-PyTorch-3D. github.com. Avail- able at: https://github.com/shijianjian/efficientne t-pytorch-3d. Accessed October 5, 2025. 31. Zeiss Humphrey Field Analyzer 3 Instructions For Use Manual. Available at: https://www.manualslib. com/manual/1548100/Zeiss-Humphrey-Field-Ana lyzer-3.html. Accessed December 2, 2025. 32. Bengtsson B, Heijl A. False-negative responses in glaucoma perimetry: indicators of patient perfor- mance or test reliability? Invest Ophthalmol Vis Sci. 2000;41:2201–2204. 33. Yohannan J, Wang J, Brown J, et al. Evidence- based criteria for assessment of visual field relia- bility. Ophthalmology. 2017;124:1612–1620. 34. Liang K, Zeger S. Longitudinal data analysis using generalized linear models. Biometrika. 1986;73:13– 22. 35. Passing H, Bablok W. A new biometrical proce- dure for testing the equality of measurements from two different analytical methods. Application of linear regression procedures for method compari- son studies in clinical chemistry, Part I. J Clin Chem Clin Biochem. 1983;21:709–720. 36. Zhang X, Francis BA, Dastiridou A, et al. Longi- tudinal and cross-sectional analyses of age effects on retinal nerve fiber layer and ganglion cell com- plex thickness by Fourier-domain OCT. Transl Vis Sci Technol. 2016;5(2):1. 37. Gardiner SK, Swanson WH, Goren D, Mans- berger SL, Demirel S. Assessment of the reliabil- ity of standard automated perimetry in regions of glaucomatous damage. Ophthalmology. 2014; 121:1359–1369.