Machine Learning to Predict Outcomes of Anti-VEGF Therapy in Neovascular Age-Related Macular Degeneration

Abstract

Objective

To develop machine learning models using OCT fluid metrics to predict long-term anti-VEGF treatment intensity and visual acuity (VA) outcomes in 2 large clinical data sets of patients with neovascular age-related macular degeneration (nAMD).

Design

Retrospective, multicenter cohort study.

Subjects

The study included a total of 2922 eyes from 2475 patients with nAMD treated at 2 tertiary centers in the United Kingdom (Belfast) and Israel (Tel Aviv Medical Center).

Methods

Longitudinal clinical data and OCT scans obtained at baseline and 6 months after treatment initiation were analyzed using a validated deep learning algorithm to extract quantitative volumetric retinal fluid measurements. Machine learning ensemble models were trained using the Belfast data set to predict total anti-VEGF injections and absolute VA at 1, 2, and 3 years from treatment initiation. External validation was performed by applying the trained models to an independent cohort to assess generalizability across health care systems.

Main Outcome Measures

Prediction accuracy for the total number of anti-VEGF injections and absolute VA at years 1 to 3 since treatment initiation, assessed by agreement with observed outcomes.

Results

In the main cohort, prediction of total anti-VEGF injections was accurate, with 99.6%, 95.5%, and 97.1% of eyes predicted within 2 injections of observed values at years 1, 2, and 3, respectively. Visual acuity predictions were within ≤0.2 logarithm of the minimum angle of resolution units (≤2 ETDRS lines) for 92.5%, 73.5%, and 76.5% of eyes at the corresponding time points. In the external cohort, model performance was reduced, with 45.0% to 53.8% of injection predictions and 26.3% to 41.1% of VA predictions meeting accuracy thresholds. Subretinal fluid at 6 months was the primary driver of injection burden predictions, whereas intraretinal fluid was most influential for VA outcomes.

Conclusions

Machine learning models based on automated fluid metrics can accurately predict treatment burden and VA outcomes within the clinical environment in which they are developed. Although performance declined with external validation, reflecting differences in treatment policies and health care systems, locally trained models remain clinically valuable. Early identification of high- and low-burden treatment trajectories may support individualized decision-making, patient counseling, and more efficient allocation of ophthalmology resources in the growing nAMD population.

Financial Disclosure(s)

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

Age-related macular degeneration (AMD) caused ∼196 million cases of blindness and vision impairment globally as of 2020.¹ Given that life expectancy is increasing, the disease burden is expected to rise to 288 million patients in 2040.² Although the neovascular AMD (nAMD) affects only ∼20% of the patients, it accounts for a substantial proportion of cases with severe central vision loss.³ Current treatment relies on intravitreal anti-VEGF injections.^4,5 Therapy is initiated once a definitive diagnosis of active nAMD is made based on retinal fluid presence on OCT.⁶ Early trials of anti-VEGF agents used fixed monthly treatment regimens.^7,8 However, current clinical practice favors more individualized strategies guided by OCT evidence of exudation, such as pro re nata (PRN) retreatment and a treat-and-extend (T&E) approach, in which the interval between injections is lengthened gradually when there are no markers of exudative disease.9, 10, 11

Patients with neovascular AMD often receive anti-VEGF injections for years, with the need for extended monitoring and treatment posing a substantial burden on both patients and health care providers. In the United Kingdom, about 40 000 new cases of nAMD are diagnosed every year, resulting in ∼1 million anti-VEGF injections administered annually with an estimated cost of >500 million Great British Pound to the National Health Service per year.¹² Patients with active nAMD have higher health care resource utilization and an average of 6.6 yearly outpatient visits.¹³ Although the treatment is effective, the health care resources are scarce, and it has become challenging for providers to supply the required demand of ophthalmology specialists. Many analyses focus on the drug costs but neglect the nondrug cost burden related to strain.¹⁴

A key aim is to predict treatment requirements and outcomes, both to manage patient expectations regarding the expected number of anti-VEGF injections and visual acuity (VA) after treatment and to improve service delivery. In the clinical setting of a T&E regimen, a patient who is expected to respond well and receive fewer injections could require fewer clinic and imaging visits compared with a patient who is predicted to have a high treatment abundance and a slower response.

Variability of patient response to anti-VEGF treatment is large, but machine learning (ML) and deep learning models have demonstrated some success in integrating the increasingly large numbers of imaging-derived and clinical features available. A recent study used clinical and imaging data from the pivotal Comparison of AMD Treatment Trials¹⁰ and predicted few (≤8) or many (≥19) injections required in a PRN regimen across 2 years based on the data from baseline to week 12 of follow-up.¹⁵ Other studies have predicted patient demand under both PRN and T&E regimens at up to 2 years, using samples of at most 600 patients.15, 16, 17, 18 There are fewer studies predicting VA after treatment, an outcome of significance for patients that can encourage adherence to intravitreal injection therapy.19, 20, 21, 22 However, these studies did not feature an external validation step, so the generalizability of the ML models is unknown, a key consideration when moving toward model deployment in the clinic.

We address this knowledge gap using data from 2 large nAMD data sets with fully automated volumetric retinal fluid analyses and long-term follow-up.²³ We developed ML models of anti-VEGF treatment response using data from 2309 nAMD eyes, based on clinical and imaging parameters from the initial 6 months of treatment to predict the number of injections and VA at up to 3 years since treatment start. Models were externally validated using 613 eyes from a different nation to assess generalizability and clinical relevance across populations.

Methods

This retrospective cohort study was conducted at Queen’s University Belfast, Belfast, UK, and at the Ophthalmology Division, Tel Aviv Medical Center (TLVMC), Tel Aviv, Israel, in a parallel manner. The research adhered to the tenets of the Declaration of Helsinki and was approved by the Institutional Review Board at TLMVC (no. 0378-20-TLV). At Queen’s University Belfast, the project was approved through the national research ethics service (research ethics committee reference no. 19/NE/0047 [amendment 1.0]). Informed consent was not required.

Study Population

Each site obtained its data from the electronic medical record (EMR) system independently, without transferring data between sites. At Queen’s University Belfast, patients with nAMD treated between May 1, 2008, and March 01, 2018 were included. Deidentified clinical data were extracted using Medisoft (Medisoft Ltd). At TLMVC, patients with nAMD treated between January 1, 2017, and October 31, 2021 were included. Data were extracted and deidentified using MDClone (MDClone).

Definition of AMD

All included patients were diagnosed with neovascular AMD based on clinical examination and imaging. Characteristic diagnostic features of nAMD on OCT findings included subretinal fluid (SRF), intraretinal fluid (IRF), or subretinal hyperreflective material. In addition, where available, fluorescein angiography demonstrating active leakage or OCT angiography showing abnormal flow consistent with macular neovascularization supported the diagnosis. The diagnosis was confirmed by the treating retina specialist at each site.

Inclusion and Exclusion Criteria

Patients included in the study were ≥50 years with a confirmed diagnosis of nAMD. Only treatment-naïve eyes, with no intravitreal anti-VEGF therapy before diagnosis, were eligible for inclusion. The included patients underwent a minimum of 3 intravitreal anti-VEGF injections within the initial 6-month period after nAMD diagnosis and had available OCT data. Both eyes of a single patient were considered eligible for inclusion if they met the criteria above. Eyes with concomitant retinal disease, such as diabetic macular edema, proliferative diabetic retinopathy, or retinal vein occlusion, were excluded from the study, as well as those with missing data (see Fig 1).

Figure 1.

Cohort construction flowchart of the Belfast-electronic medical record data set. VA = visual acuity.

Data

Clinical data included patient characteristics (age, gender, and treated eye). Visual acuity values recorded as Snellen entries were converted to logarithm of the minimum angle of resolution (logMAR) scale for statistical analysis and then categorized into 4 groups: good VA <0.3 logMAR; moderate VA ≥0.3 to <0.6; partially sighted VA ≥0.6 to <1.0; and low vision VA ≥1.0. Intravitreal injections were recorded at each visit. A uniform corporate data dictionary was defined to ensure harmonization of the data set at both study sites.²³ Imaging data were obtained using the Heidelberg Spectralis OCT (Heidelberg Engineering), anonymized and stored in an isolated, secure environment in the proprietary Heidelberg E2E file format.

Deep Learning OCT Analysis

All OCT volume scans were analyzed using a deep learning and ML software algorithm (Notal OCT Analyzer [NOA], Notal Vision Ltd), which segments OCT B-scans in an automated manner and quantifies anatomic retinal fluid parameters. The NOA algorithm, its application in nAMD, and its diagnostic accuracy have been previously validated on Cirrus and Spectralis systems.²⁴ The algorithm performs an autonomous fluid analysis from which we extracted the 42 parameters listed in Table S1 (available at https://www.ophthalmologyscience.org), including fluid parameters such as SRF, IRF, and total retinal fluid volumes (nL); retinal volume (nL); volume of retinal pigment epithelium irregularities (nL); volume of retinal pigment epithelial detachment (nL), and more. Each parameter is provided for the 9 ETDRS standard fields, allowing separate analysis of the central 1-mm area and the 3- and 6-mm areas, as well as the total scan area.

Outcome Measures

We predicted 2 sets of outcomes in treatment naïve eyes during the first, second, and third years of treatment: (1) total number of anti-VEGF injections and (2) absolute VA. Predictions were made first using the Belfast-EMR data set, followed by an external validation using the TLVMC data set.

Prediction Models

Inputs to our prediction model comprised fluid measurements provided by the NOA algorithm before the first anti-VEGF injection (at baseline) and at the closest clinic visit 6 months after treatment initiation. These time points were selected because they reflect disease severity at baseline, an important determinant of long-term outcomes, and provide insight into the initial response to treatment at the end of the intensive loading phase. The NOA algorithm provides multiple quantitative fluid measurements for each OCT volume scan (Table S1), and we had no a priori expectations regarding which variables would be most important for prediction or about the functional form of associations among variables. Therefore, we used a flexible ML approach, fitting a library of models and combining their outcomes to make the final predictions. This ensemble method, called SuperLearning, yields a final model that performs at least as well as the best of the candidate models specified, and often substantially better.²⁵ Models in our library included least absolute shrinkage and selection operator and ridge regression, polynomial splines, neural networks, XGBoost, and random forests—a diverse set able to model nonlinear associations and interactions between variables. To reduce the risk of overfitting, we used 10-fold cross-validation. Where measurements were available for both eyes of a single patient, both eyes were assigned to the same cross-validation split to prevent information leakage from inflating model performance. For both tasks, the loss function was mean squared error. Separate ensembles were fitted for each task and at each time point.

We investigated the generalizability of our models by performing a full external validation, developing the models using the Belfast-EMR data set and then applying the trained model to the independent TLVMC data set. Model performance was summarized using descriptive statistics and linear calibration curves. Calibration in the large was assessed, along with weak calibration (linear curves). Moderate calibration (flexible curves) was not assessed.

Interpretability

We investigated which of the fluid measurements had the greatest influence on predictions by analyzing Shapley Additive Explanations (SHAP) values, an interpretability technique based on game theories of apportioning reward. In every modeling task, the prediction for each eye is decomposed into contributions from each of the predictor variables. In this study’s modeling tasks, a positive SHAP value would indicate that the observed value of the predictor (either positive or negative) increased the predicted number of injections. We ranked variables by the mean of their absolute SHAP values when predicting outcomes at the end of 3 years, the most impactful variables being those with the greatest mean. Ensemble models were fitted using the sl3 package, and SHAP values were calculated using the fastshap package in R (version 4.3.3).^26,27

Results

The prediction model used the Belfast-EMR data set, which was initially comprised of 7370 eyes of 5618 patients with nAMD. After application of the inclusion and exclusion criteria, 2309 eyes of 1949 patients were included in the study (Fig 1). The TLVMC external validation data set was similarly constructed and contained 613 eyes of 526 patients. The median age of patients was 79.1 years in Belfast-EMR and 80.6 years in TLVMC, with 66.8% and 54.6% female patients, respectively. The patients’ demographic and clinical characteristics at baseline, alongside their VA and OCT-derived fluid measurements, are presented in Table 1. Median baseline VA was similar between the data sets (0.50 logMAR in Belfast-EMR vs. 0.48 logMAR in TLVMC). However, their distribution differed, with the Belfast-EMR data set having more patients with moderate and partially sighted VA (31.7% and 29.1%), and the TLVMC data set showed a higher proportion with good VA (36.9%) and low vision (23%). During the follow-up period, Belfast-EMR eyes had more VA (median 25 vs. 11) and OCT-derived fluid measurements (median 15 vs. 9) per eye compared with TLVMC. However, the TLVMC eyes received more injections (median 17 vs. 11).

Table 1.

Demographic and Clinical Characteristics of Belfast-EMR and TLVMC Cohorts at Baseline and at the 3-Year Follow-Up Period

Variable	Belfast-EMR	TLVMC
Patients, n	1949	526
Eyes, n	2309	613
Baseline characteristics
Age, yr, mean (range)	79.1 (51.3–99.7)	80.6 (50.1–101)
Gender, female, n (%)	1543 (66.8%)	353 (54.6%)
VA, logMAR, median (range)	0.5 (0–2.3)	0.48 (0–2.3)
Good VA (<0.3), n (%)	574 (24.9%)	223 (36.9%)
Moderate (≥0.3 to <0.6), n (%)	733 (31.7%)	119 (19.7%)
Partially sighted (≥0.6 to <1.0), n (%)	672 (29.1%)	124 (20.5%)
Low vision (≥1.0), n (%)	330 (14.3%)	139 (23%)
3-yrs follow-up period
Total anti-VEGF injections, n	27 341	10 102
Injections per eye, median (IQR)	11 (6–16)	17 (10–23)
Total VA measurements	55 564	6666
VA measurements per eye, median (range)	25 (4–61)	11 (1–38)
Total OCT fluid measurements	33 286	5569
Fluid measurements per eye, median (range)	15 (3–59)	9 (2–28)

EMR = electronic medical record; IQR = interquartile range; logMAR = logarithm of the minimum angle of resolution; n = sample size; TLVMC = Tel Aviv Medical Center; VA = visual acuity.

Total Number of Anti-VEGF Injections

Using retinal fluid measurements from baseline and 6-month time points, the total number of anti-VEGF injections was predicted with high accuracy in the Belfast-EMR data set. Overall, 99.6% of predictions were within 2 injections of the number observed in year 1, 95.5% in year 2, and 97.1% in year 3 (Table 2). Calibration slopes ranged between 1.52 and 1.64, indicating a tendency to slightly overestimate the number of injections for eyes with lower treatment needs and underestimate them for eyes with higher treatment needs (Fig 2).

Table 2.

Performance of Prediction Models for the Number of Anti-VEGF Injections and Visual Acuity

Predicted Outcome
Total Injections	Year	Belfast-EMR		TLVMC
		Mean Absolute Error	Within 2 Injections	Mean Absolute Error	Within 2 Injections
	1	0.74	99.6%	2.75	45.0%
	2	1.10	95.5%	2.08	39.0%
	3	1.01	97.1%	3.11	53.8%
Visual Acuity			Within 2 Lines		Within 2 Lines
	1	0.09	92.5%	0.38	41.1%
	2	0.15	73.5%	0.40	33.1%
	3	0.14	76.5%	0.41	26.3%

EMR = electronic medical record; TLVMC = Tel Aviv Medical Center.

Figure 2.

Predicted number of injections versus number of injections received by treatment year for Belfast-electronic medical record data set. Solid lines indicate calibration slopes (blue = fitted, black = perfect).

The most influential variable for predicting 3-year outcomes was SRF volume at 6 months, with 9 of the top 10 most important predictors derived from the 6-month time point, the exception being total retinal volume measured at baseline (Fig S1, available at https://www.ophthalmologyscience.org). Subretinal fluid and retinal pigment epithelium irregularity measurements ranked higher in prediction importance than IRF measurements. The presence of any SRF at 6 months increased the predicted number of injections across all 3 years.

In the TLVMC external validation set, performance was lower with 45.0%, 39.0%, and 53.8% of predictions falling within 2 injections of the observed values for years 1, 2, and 3, respectively. The model underpredicted the number of injections compared with the number observed for each of the 3 years, with calibration intercepts >1 (Fig 3). Calibration slopes were relatively close to 1, ranging from 0.80 to 1.06, indicating weak discrimination by the number of predicted injections. The ensembles were heavily weighted toward predictions from the Random Forest model, with polynomial splines and XGBoost models featuring to a lesser degree (Table S2, available at https://www.ophthalmologyscience.org).

Figure 3.

Predicted number of injections versus number of injections received by treatment year for Tel Aviv Medical Center data set. Solid lines indicate calibration slopes (blue = fitted, black = perfect).

Visual Acuity

In the Belfast-EMR data set, the model predicted VA at the end of each year with high accuracy (Table 2). Predictions were within ≤0.2 logMAR units (≤2 ETDRS lines) for 92.5%, 73.5%, and 76.5% of eyes in years 1, 2, and 3, respectively. Calibration slopes ranged between 1.30 and 1.49, indicating that predictions were generally not sufficiently extreme: eyes predicted to have good VA tended to perform better than anticipated, whereas among those with a prediction of poor VA at the other end of the distribution, observed VA was worse than predicted (Fig 4).

Figure 4.

Predicted versus observed visual acuity at the end of each treatment year for Belfast-electronic medical record data set. Solid lines indicate calibration slopes (blue = fitted, black = perfect). VA = visual acuity.

In the TLVMC external validation set, the model’s performance was lower (Table 2). In year 1, 41.1% of eyes were within ≤0.2 logMAR units of observed values, with the rate standing at 33.1% and 26.3% for years 2 and 3, respectively. Calibration slopes ranged between 1.54 and 1.68, again suggesting insufficiently extreme predictions (Fig S2, available at https://www.ophthalmologyscience.org). The ensembles were dominated by the random forest model, with smaller contributions from the XGBoost and linear models (Table S3, available at https://www.ophthalmologyscience.org).

Intraretinal fluid and retinal volumes were the predominant model drivers for predicting VA at 3 years, accounting for 9 of the top 10 prediction variables (Fig S3, available at https://www.ophthalmologyscience.org). The most influential variable was IRF volume in the temporal 3-mm area at baseline, followed by total IRF volume across all regions at both baseline and 6 months. The presence of IRF, particularly when present at both baseline and 6 months, shifted predictions toward substantially worse VA. Small retinal volumes in the central and inner inferior, nasal, and superior regions at 6 months were associated with poorer VA as well.

Discussion

We developed ML-based models to predict long-term anti-VEGF treatment intensity and VA outcomes in patients with nAMD using only quantitative OCT-derived fluid metrics obtained at baseline and at the first 6-month time point of treatment. We demonstrated that anatomic biomarkers at baseline and early treatment response can support more personalized patient counseling and offer prognostic insight to their future treatment needs, thereby contributing to individualized treatment planning. We used 2 large real-world data sets and an ensemble learning framework to flexibly predict the relationships between fluid measurements and outcomes without prior knowledge of their functional form. Tree-based models that can capture thresholds and interactions among variables were most successful in prediction driving. For predicting the number of injections, measurements of SRF were most important, indicating that SRF presence was a key determinant of treatment. This finding aligns with the results reported by Bogunović et al¹⁶ and Tong et al²⁸ who similarly found that the SRF volume, particularly in the central 3-mm circle, had the highest predictive values. Intraretinal fluid volumes were most influential for predicting VA, consistent with previous reports showing that persistent IRF was strongly associated with poorer visual outcomes.^19,29, 30, 31, 32 Although the models demonstrated high performance when trained and cross-validated within a single center, performance was reduced in the external validation cohort.

Prediction accuracy of our model for the number of injections was slightly higher than reported in previous studies,^15,17,18 and for VA, model performance was similar to other reports.^19,20,22 These generally relied on smaller data sets in comparison to the one used in the current study, with 2309 eyes in the training set and 613 eyes in the external validation cohort.²³ Data from randomized controlled trials (RCTs) offer several advantages for model development, including standardized treatment protocol design, high-quality clinical and imaging data, consistency, and closely monitored follow-up schedules. Chandra and Ying¹⁵ performed a secondary analysis of the Comparison of AMD Treatment Trials study data to predict treatment demand of few (≤8) or many (≥19) injections and the total number of injections over 2 years. Their model incorporated demographic variables together with OCT features from baseline through week 12. Among 493 participants, the model achieved areas under the curves (AUCs) of 0.77 and 0.82 for predicting few and many injections, respectively. Bogunović et al¹⁶ analyzed data from the HARBOR study (317 patients),¹¹ to predict low (≤5), medium, and high (≥16) injection requirements in a PRN regimen. They showed an AUC of 0.70 and 0.77 for prediction in the low- and high-intensity groups, respectively. Although these data sets are well curated, they lack the variability and complexity of real-world populations and are less applicable for generalizability.

Real-world cohorts provide important advantages for predictive modeling, including larger sample sizes and greater relevance to routine clinical practice. They capture the diversity of patient populations, variation in treatment regimens, and anti-VEGF agents. At the same time, they introduce challenges in the form of inconsistent follow-up intervals, nonattendance, and missing data. Gallardo et al¹⁷ included 379 eyes (340 patients) with nAMD from real-world setting, to predict low (≤5) and high (≥10) injection need across 1 year in a T&E setting. The nAMD model yielded mean AUCs of 0.79 for both groups. Pfau et al¹⁸ included 99 eyes from 96 patients treated with PRN or T&E regimens using an automated OCT-based pipeline to predict injection number and received a mean absolute error (MAE) of 2.60 injections/yr, similar to the performance of our model within the external validation data set (MAE = 2.75).

Although treatment decisions in nAMD management are predominantly guided by OCT-based findings, many health care systems still apply VA thresholds for treatment eligibility, and patients often prioritize anticipated visual outcomes when discussing prognosis. For these reasons, we also aimed to predict VA outcomes. We predicted VA outcomes with reasonable accuracy, with most predictions falling within 0.2 logMAR units (≤2 ETDRS lines) of observed VA values. Two previously published ML studies predicting VA outcomes relied on RCT data. Chandra and Ying²² used the Comparison of AMD Treatment Trials data set in a cohort of 1029 participants, achieving an MAE of 8.8 to 10.2 letters for predicting final VA at 2 years. Schmidt-Erfurth et al¹⁹ analyzed the HARBOR RCT data using VA and OCT biomarkers to predict 12-month VA and achieved an accuracy of R² = 0.70, with IRF as the most relevant predictive biomarker. It is noteworthy that VA in real-world data sets reflects recorded VA rather than standardized best-corrected VA, and it is possible that the use of best-corrected VA, available in RCT settings, might have improved model performance. Two additional studies used real-world data. In a cohort of 456 patients, ML algorithms used OCT metrics and clinical parameters to predict VA at 1 year and found the difference between prediction and observation to be MAE 0.16 to 0.22 logMAR units.²⁰ An additional real-world study from Taiwan (698 patients) applied deep learning models and presented a high discriminatory ability (AUC = 0.989) in predicting VA improvement of ≥2 lines.²¹

External validation of predictive models is a crucial step to being clinically applicable and relevant.³³ To our knowledge, external validation on nAMD data from a different country has not been performed in prior ML studies. As expected, model performance of the Belfast-trained models for both tasks was considerably lower when externally validated using TLVMC data; models systematically underpredicted the number of injections when applied in the external cohort. In regard to visual outcomes, eyes predicted to have good VA tended to perform better than anticipated, whereas those predicted to have poor VA performed worse. Hence, these findings suggest that although the models function well within the site they were trained on, their transferability to other cohorts is limited.³⁴ A possible explanation lies in substantial differences in treatment policies and care pathways between countries. In Israel, treatment is commonly initiated with bevacizumab as a first-line agent, whereas in the United Kingdom it is considered off-label, with ranibizumab or aflibercept used as approved first-line agents during the period of data acquisition. Furthermore, TLVMC patients had fewer VA and OCT monitoring visits but received more injections per year, reflecting different treatment regimens in clinical routine.³⁵ The Israeli system does not set VA restrictions for treatment eligibility, whereas in the United Kingdom, the National Institute for Health and Care Excellence guidelines recommend ranibizumab injections to the eyes with VA between 6/12 and 6/96 (equivalent to 0.3–1.2 logMAR).³⁶ The differences above represent a methodological and health system–driven difference in nAMD treatment, rather than a biological difference in disease activity, yet they could significantly influence the treatment dynamics reflected in the data. In addition to cross-country health care system differences, temporal variation in data extraction periods may have contributed to reduced model performance as well. The TLVMC validation cohort reflects more recent clinical practice (2017–2021) compared with the Belfast data set (2008–2018), during which anti-VEGF treatment evolved, with wider adoption of T&E regimens and shifts in preferred agent choices. These changes in practice over time may have influenced injection patterns and possibly contributed to the model’s reduced transportability.

Given these challenges in cross-site transferability, it is important to consider the value of models developed specifically for a local clinical environment. Locally trained and deployed models may offer substantial clinical value by accurately predicting treatment trajectories for the center’s or country’s own patient population. Model performance can be regularly evaluated postdeployment. A dynamic modeling framework could adapt to sudden changes, such as the launch of a new drug or a change in dosing regimens. This approach, combined with model explanation techniques, could produce robust single-site models that may be extended to include procedural data captured in a hospital EMR to improve performance. Eventually, the model’s predictions can support service planning, stratify patients into high and low intensity treatment trajectories, and optimize resource allocation. Significant impact can be achieved through site-specific models that are tailored to local protocols without necessarily requiring generalization across different health systems.

The study has several notable strengths. With its large real-world patient cohort, the study data reflect routine clinical practice and enhance the applicability of our findings to real-world management of nAMD. The use of a fully automated, volumetric OCT algorithm offers a scalable and universally implementable approach. Furthermore, the incorporation of a binational system allowed us to address the external validity question that has not been evaluated in previous studies in the field. Nevertheless, limitations should be acknowledged. Because of the study’s retrospective design and the inherent heterogeneity of EMR-derived data, some potentially important variables were not available, such as macular neovascularization subtype classification, specific anti-VEGF agent administered, and systemic and ocular comorbidities that could affect visual outcomes. Although the study design allowed for a scalable analysis, the incorporation of additional baseline clinical variables and disease characteristics may further improve predictive performance. Notably, the macular neovascularization subtype may impact disease course and treatment intensity requirements.³⁷ Future predictive models could benefit from lesion subtype stratification, integration of additional clinical information, and potentially genomic data, which are becoming more accessible, alongside fluid metrics to improve predictive accuracy and transportability.

Looking ahead, the accelerating integration of artificial intelligence (AI) into clinical practice is likely to further expand the role of ML in driving clinical decisions. Our approach provides a practical roadmap for single-site modeling of nAMD outcomes, representing an important step to real-world clinical translation. Moving toward clinical implementation, future work should ensure adherence to the principles outlined in the FUTURE-AI framework for deployable AI in health care and address regulatory considerations relevant to model deployment.³⁸ These tools may be particularly valuable in a highly prevalent condition, such as nAMD, where the clinical and economic burden is expected to increase with population aging. More efficient resource allocation, guided by data-driven predictions, has the potential to reduce both patient and health care system strain. The availability of AI-driven platforms together with emerging technologies such as home OCT may further enhance the clinical decision-making process and enable more personalized, proactive care. Leveraging these technologies to support clinicians and provide patients with clearer insights regarding their condition represents a promising path toward more informed management of nAMD.

Manuscript no. XOPS-D-26-00117R1.