Artificial Intelligence for Predicting Postoperative Outcomes in Anterior Segment and Ocular Adnexa Surgery: A Systematic Review

Abstract

Topic

Artificial intelligence (AI) is increasingly applied to support decision-making in ophthalmology. This review evaluates the ability of AI to predict postoperative outcomes after anterior segment and ocular adnexal procedures.

Clinical Relevance

The prognostic application of AI algorithms, capable of handling multimodal, complex data, and modeling nonlinear relationships, may enhance perioperative patient counseling, surgical planning, and outcomes.

Methods

A comprehensive search was conducted across MEDLINE, Embase, the Cochrane Library, Compendex, Web of Science, Scopus, ProQuest Dissertations and Theses, and gray literature. Studies that investigated only intraocular lens power calculation and refractive outcomes were excluded. Methodological quality was assessed using the Prediction Model Risk of Bias Assessment Tool.

Results

Twenty-eight studies were included (n = 2 871 898 eyes). Most procedures were corneal and ocular surface (50%), followed by glaucoma (25%) and cataract/oculoplastic procedures (25%). Tree-based algorithms (57%) dominated the representation layer, followed by deep learning algorithms (43%). Fourteen studies examined how AI performance compared with traditional prediction paradigms; in 13 (93%), AI outperformed traditional methods. Areas under the curve ranged from 0.33 to 0.99, reflecting substantial variation in outcome definitions and model design. The Prediction Model Risk of Bias Assessment Tool assessment indicated that 75% of studies had high risk of bias, primarily due to insufficient reporting of model transparency and calibration metrics. Only 11% of studies incorporated external validation.

Conclusions

Artificial intelligence models are promising in predicting anterior segment and adnexal ophthalmic surgical outcomes. While these models generally outperform conventional methods, methodological limitations persist. Future studies should incorporate external validation, standardized reporting, and assessment of workflow integration.

Financial Disclosure(s)

The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.

Predicting complications and functional outcomes after ophthalmic surgery would enhance perioperative management, planning, and patient counseling, and ultimately patient care.¹ As the number of ophthalmic procedures worldwide continues to increase, this expanding surgical load increases the risk of complications, exacerbated by the growing proportion of elderly patients, a population with greater medical complexity and comorbidities, which can further affect postsurgical recovery.^2,3 Moreover, surgeries with complicated outcomes and recoveries increase health systems costs by an estimated 119%–172% in comparison to treatments devoid of complications.⁴ As demonstrated, complex postsurgical recovery leads to downstream systemic effects that extend beyond the individual patient's clinical course. However, accurate prediction of postoperative outcomes remains a challenge due to limitations inherent in traditional statistical models, the heterogeneity of procedures examined, and the variability in how these outcomes are defined.

In current practice, preoperative visual acuity, age, preexisting ocular comorbidities, and systemic health status are among the features typically used to predict postoperative outcomes.5, 6, 7 However, integrating these features into conventional statistical tools such as logistic regression or Cox proportional hazards models can raise issues due to their assumptions of linear relationships among variables and reliance on small predictor sets. These constraints can thereby obscure the complex interactions among demographic, clinical, imaging, and surgical parameters that meaningfully influence postoperative outcomes.⁸

By modeling nonlinear relationships and integrating high-dimensional data, artificial intelligence (AI) has the potential to address these limitations. For instance, machine learning (ML) algorithms can incorporate a greater range of variables to uncover complex interactions that may not be appreciated by traditional analyses or clinician judgment.^8,9 Deep learning (DL), a subset of ML which involves multilayered neural networks, can further delineate highly discriminative features within large data sets involving imaging and clinical parameters to further enhance prediction of postoperative outcomes.¹⁰

To date, AI has been applied to predict a host of ophthalmic surgical outcomes. These include the prediction of graft survival after corneal transplantation, outcomes of laser vision correction, surgical success in glaucoma procedures, cataract surgery complications, and aesthetic outcomes after oculoplastic interventions.11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38 These studies use diverse input predictors, including patient demographics, biometric data, and imaging parameters, to predict these outcomes. While these results are encouraging, the literature remains vastly heterogeneous, with variability in data quality, sample size, and validation methods.11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38 Outcome definitions also vary substantially across studies. Importantly, external validation or prospective evaluation methods to assess these AI models are rarely incorporated into study design, limiting their translation and integration into clinical practice. Recently, a systematic review and meta-analysis examined the utilization of AI algorithms to predict postoperative outcomes in vitreoretinal surgery; however, a synthesis of these applications in anterior segment and ocular adnexa procedures has not been conducted to date.³⁹

The objective of this systematic review is to therefore evaluate the predictive performance, methodological rigor, and clinical translation readiness of AI models in anterior segment and adnexal surgeries. This review also aims to compare the performance of these algorithms to conventional approaches such as logistic regression or clinician judgment. By consolidating these findings and synthesizing the available evidence, future considerations and research directions to support the development of robust models and facilitate their clinical translation in ophthalmic surgery will be elucidated.

Methods

Following registration on the International Prospective Register of Systematic Reviews (CRD420251065418), this review was conducted according to the guidelines outlined by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses. This study was conducted in accordance with the Declaration of Helsinki. Institutional Review Board approval (from the University of Calgary) and patient informed consent were not required, as all data were derived from previously published studies.

Eligibility Criteria

Studies were eligible for inclusion if they involved patients undergoing anterior segment or ocular adnexa ophthalmic surgery and involved the application of AI to predict postoperative outcomes. Eligible surgical domains included cataract surgery, glaucoma filtration surgery, corneal transplantation, and oculoplastic procedures. Studies employing internal and external validation techniques were deemed eligible, provided they reported quantitative predictive performance metrics, including accuracy, area under the receiver operating characteristic curve (AUC), C-statistic, sensitivity, specificity, and mean absolute error. Comparator models, such as traditional statistical methods or clinician-based predictions, were included when available, but were not mandatory. Eligible study designs encompassed retrospective and prospective cohort studies, and randomized and nonrandomized clinical trials. Reviews, commentaries, editorials, and case reports were excluded. Studies were also excluded if they evaluated only intraocular lens power calculation formulas or refractive outcomes (e.g., postoperative spherical equivalent or residual refractive error).

Search Strategy

A comprehensive search was conducted across MEDLINE, Embase, the Cochrane Database of Systematic Reviews, the Cochrane Central Register of Controlled Trials (CENTRAL), IEEE Xplore, Compendex, Web of Science, Scopus, and ProQuest Dissertations and Theses. Gray literature was reviewed through Google Scholar. Searches covered all records from database inception to the date of the final search (June 3, 2025) and were restricted to English-language publications.

Search strategies combined terms related to the following: (1) anterior segment and ocular adnexal surgery (e.g., cataract surgery, keratoplasty, trabeculectomy, oculoplastic procedures); (2) AI and ML (e.g., neural networks, DL); and (3) postoperative outcomes (e.g., prognosis, risk stratification). The full search strategy is available in Supplemental Table 1, available at www.ophthalmologyscience.org.

Two independent reviewers (M.A. and L.Y.) screened titles and abstracts using a standardized screening form. Discrepancies were resolved through discussion with a third supervising reviewer (A.A.A.), who was masked to the initial screening decisions.

Data Extraction

Data were extracted using a standardized form to ensure consistency across studies. One reviewer (M.A.) performed initial data extraction, and a second independent reviewer (L.Y.) verified all extracted data independently by re-extracting each field and comparing entries against the primary extraction. Discrepancies identified between reviewers were first addressed through direct discussion and consensus. Any disagreements that could not be resolved by the 2 primary reviewers were escalated to a third reviewer for adjudication (A.A.A.). Extracted information included the studies’ characteristics, such as the article title, authors, year of publication, country, study design, patient and eye sample sizes, and mean participant age. The study’s methodological details were also retrieved, including inclusion and exclusion criteria, the type of validation employed (internal or external), and the duration of postoperative follow-up. Moreover, AI model characteristics were captured, including the algorithm family, input features, comparator methods (where applicable), and predicted outcomes. The performance metrics reported by included articles were also extracted, including accuracy, AUC, C-statistic, sensitivity, specificity, and mean absolute error.

Risk of Bias Assessment

Methodological quality and applicability were evaluated using the Prediction Model Risk of Bias Assessment Tool for AI, an extension of the original Prediction Model Risk of Bias Assessment Tool which assesses the risk of bias and applicability of prognostic models incorporating ML or AI algorithms.⁴⁰ This tool is structured around 4 domains (participants and data sources, predictors, outcomes, and analysis) with separate assessment pathways for model development and model evaluation. Moreover, applicability is evaluated across the first 3 domains (participants and data sources, predictors, and outcomes).⁴¹ Studies judged to have a high risk of bias in any single domain were classified as overall high risk. When assessment within a domain was unclear, but no high-risk evidence was identified, the study was classified as “unclear risk.” Studies with ≥1 domain rated as having low or unclear applicability were considered to have low or unclear overall applicability. Two reviewers (M.A. and L.Y.) conducted assessments independently. Discrepancies were resolved through discussion with a third supervising reviewer (A.A.A.), who was masked to the initial decisions.

Data Synthesis

Given the heterogeneity of model architectures, this synthesis was exploratory and did not represent a formal meta-analysis. In light of this limitation, a thorough narrative synthesis was undertaken to summarize and integrate the findings. Results were qualitatively grouped by surgical domain (e.g., cornea, glaucoma, cataract, or adnexa) and AI modeling family (e.g., tree-based, neural network, regression-based).

Results

Study Selection and Characteristics

The initial database search yielded 2575 unique records. From these findings, only 60 articles were deemed eligible for full-text review, as they were identified as primary studies involving the application of AI in postoperative prediction for anterior segment and ocular adnexa surgery and reported on model performance metrics. After completion of the screening process, 28 studies were found to meet the inclusion criteria (Fig 1). This entailed 2 871 898 eyes from 2 869 845 participants across 14 countries (Supplementary Fig S1, available at www.ophthalmologyscience.org). The largest sample size was reported in the study by Triepels et al,³⁶ which involved 2 853 376 eyes. The United States and India were the most frequently represented study origins (each n = 4, 14%), followed by Italy (n = 3, 11%). Most studies were retrospective in design (n = 23, 82%), with only 3 of prospective design (10%). The recency of publications was notable, with 82% (n = 23) of articles published in 2022 or later.

Figure 1.

PRISMA flow diagram. PRISMA = Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Sample sizes varied substantially, from 44¹⁵ eyes to 2.85 million³⁰ eyes (Table 1).^21,36 Models reporting the highest AUCs (≥0.90) tended to originate from smaller data sets (Fig 2). Predictor counts ranged from 1 to 432. While the majority of studies incorporate patient factors exclusively such as corneal thickness or preoperative intraocular pressure (IOP), Muijzer et al, Mohammadi et al, and Lanza et al^26,29,30 also involved surgeon-specific factors such as level of training.

Table 1.

Characteristics of Included Studies for Review

Author (Year)	Country	Sample Size (Eyes)	Surgical Domain	Procedure(s)	Mean Age (Years)	% Female	Follow-Up Duration	Study Design
Karaca, Emine (2025)	Turkey	209	Cornea and External Disease	DMEK	N/A	N/A	1 year	Retrospective cohort
Kim, Min Seok (2025)	South Korea	106	Cornea and External Disease	DMEK	59.8	49.1	21.2 months	Retrospective cohort
Kundu, Gairik (2025)	India	250	Cornea and External Disease	DALK	28.0	46	70.8 months	Retrospective cohort
Yang, Che-Ning (2025)	Taiwan	543	Cataract and Anterior Segment	Cataract Surgery	69.2	64.3	3.0-6.0 months	Retrospective cohort
Barry, Samuel (2024)	United States of America	2398	Glaucoma	Trabulectomy, ExPress shunts, tube shunts, MIGS, CPC	72.5	47.4	3.0+ months	Retrospective cohort
Birla, Shweta (2024)	India	207	Glaucoma	Trabulectomy	N/A	33.0	5 years	Retrospective cohort
Guzun, Olga (2024)	Ukraine	127	Glaucoma	CPC	65.0	67.0	1 year	Retrospective cohort
Lee, Seung (2024)	South Korea	153	Glaucoma	Ahmed valve implantation	61.5	34.5	1 year	Retrospective cohort
Ling, Clarissa (2024)	Singapore	1687	Cornea and external disease	PK, DALK	51.8	45.9	10 years	Prospective cohort
Lin, Wei-Chun (2024)	United States of America	1540	Glaucoma	Trabeculectomy	63.6	57	1 year	Retrospective cohort
Agnifili, Luca (2023)	Italy	102	Glaucoma	Trabulectomy	N/A	40.1	1 year	Prospective cohort
Francis, Mathew (2023)	India, Italy	529	Cornea and external disease	SMILE, LASIK, PRK	26.0	54.9	7.1 months	Retrospective cohort
Hayashi, Takahiko (2023)	Germany	312	Cornea and external disease	DMEK	69.2	51.6	4 weeks	Retrospective cohort
Joseph, Naomi (2023)	The Netherlands	44	Cornea and external disease	Keratoplasty	N/A	N/A	1-72 months	Retrospective cohort
Patefield, Alastair (2023)	United Kingdom	74	Cornea and external disease	DMEK	68.5	62.2	N/A	Retrospective cohort
Triepels, Ron (2023)	International (mainly Europe)	2 853 376	Cataract and anterior segment	Cataract surgery	73.9	58.7	N/A	Retrospective cohort
Ang, Marcus (2022)	Singapore	1335	Cornea and external disease	PK, DSAEK	68.3	52.4	10 years	Prospective cohort
Banna, Hasan (2022)	United States of America	230	Glaucoma	Trabulectomy	69.2	41.3	1 year	Retrospective cohort
Muijzer, M.B. (2022)	The Netherlands	3647	Cornea and external disease	DSEK, DMEK	73.0	N/A	1 year	Retrospective cohort
Patel, Sanjay (2022)	United States of America	48	Cornea and external disease	DMEK	68.0	59.0	6 months	Prospective cohort
Santhiago, Marcony (2022)	Brazil	339	Cornea and external disease	LASIK	31.5	44.7	5 years	Retrospective case-control
Sun, Yiming (2022)	China	450	Oculoplastic and orbital surgery	Blepharoptosis surgery	7.9	35.6	115.7 days	Retrospective cohort
Wang, Jinhong (2022)	China	1393	Cataract and anterior segment	Cataract surgery	N/A	N/A	1 month	Retrospective cohort
Hayashi, Takahiko (2021)	Germany	46	Cornea and external disease	DALK	48.1	23.9	N/A	Retrospective cohort
Lanza, Michele (2021)	Italy	1392	Cataract and anterior segment	Cataract surgery	1392	71.3	N/A	Retrospective cohort
Yoo, Tae Keun (2020)	South Korea	218	Oculoplastic and orbital surgery	Orbital decompression surgery for thyroid eye disease	N/A	N/A	N/A	Retrospective cohort
Gupta, Manish (2016)	India	791	Cornea and external disease	LASIK	N/A	53.2	1 month	Retrospective cohort
Mohammadi, Seyed-Farzad (2012)	Iran	352	Cataract and anterior segment	Phacoemulsification for age-related cataract	66.1	48.6	2 years	Cross-sectional

CPC = cyclophotocoagulation; DALK = deep anterior lamellar keratoplasty; DMEK = Descemet membrane endothelial keratoplasty; DSAEK = Descemet's stripping automated endothelial keratoplasty; DSEK = Descemet stripping endothelial keratoplasty; LASIK = laser-assisted in situ keratomileusis; MIGS = microinvasive glaucoma surgery; PK = penetrating keratoplasty; PRK = photorefractive keratectomy; SMILE = small incision lenticule extraction.

Figure 2.

Bubble plot of study size versus AI performance (AUC). AI = artificial intelligence; AUC = area under the curve.

Surgical Domains and Procedures

Studies investigating the application of AI-based models in cornea and ocular surface disease were the most common (n = 14, 50%), followed by glaucoma (n = 7, 25%), cataract/anterior segment (n = 5, 18%), and oculoplastic/orbital surgery (n = 2, 7%). Pertaining to individual procedures, corneal transplantation (n = 11, 39%), cataract extraction (n = 5, 18%), and trabeculectomy (n = 5, 18%) were the most frequently evaluated. Among the investigated outcomes, corneal graft failure (n = 4, 14%) and glaucoma filtration surgery failure (n = 6, 21%) were the most prevalent.

AI Model Families and Comparators

Tree-based ensemble algorithms (random forests, gradient boosting, and Extreme Gradient Boosting) accounted for 57% (n = 16) of all models in the included studies. Following these algorithms, neural networks (convolutional neural networks, multilayer perceptrons, and generative adversarial networks) accounted for 43% (n = 12) (Table 2). Artificial intelligence models demonstrated superior discriminative performance in 13 of 14 studies that involved direct comparisons with conventional statistical methods, typically manifested in higher AUC or accuracy (Fig 3).

Table 2.

AI Models, Comparators, Outcomes, and Validation Across Included Studies

Author (Year)	Surgical Domain	AI Model Family	Specific Model(s)	Comparator	Outcome Predicted	Did AI Outperform the Comparator?	Validation
Karaca, Emine (2025)	Cornea and external disease	Neural network	1-ANN	None	Early corneal graft failure	N/A	Internal (hold-out)
Kim, Min Seok (2025)	Cornea and external disease	Tree-based (ensemble)	XGBoost, random Forest, CatBoost (best)	Logistic regression	Postoperative ocular hypertension	Yes	Internal (K-fold)
Kundu, Gairik (2025)	Cornea and external disease	Tree-based	Random Forest	None	Loss of corneal graft clarity, scarring at graft–host interface	N/A	Internal (hold-out)
Yang, Che-Ning (2025)	Cataract and anterior segment	Neural network	Xception	None	Postoperative visual acuity	N/A	Internal (K-fold)
Barry, Samuel (2024)	Glaucoma	Tree-based	Random Forest	Logistic regression	Trabulectomy, ExPress shunts, tube shunts, MIGS, CPC failure	Yes	Internal (K-fold)
Birla, Shweta (2024)	Glaucoma	Tree-based	trees.FT	Logistic regression	Trabulectomy failure	Yes	Internal (K-fold), external
Guzun, Olga (2024)	Glaucoma	Neural network	MLP	None	CPC failure	N/A	Internal (hold-out)
Lee, Seung (2024)	Glaucoma	Tree-based	XGBoost	Logistic regression	Ahmed valve implant failure	Yes	Internal (K-fold)
Ling, Clarissa (2024)	Cornea and external disease	Tree-based	Random survival forest	Cox proportional hazards	Postoperative best corrected visual acuity	Yes	Internal (bootstrap)
Lin, Wei-Chun (2024)	Glaucoma	Neural network	Transformer-based multimodal	None	Trabulectomy failure	N/A	Internal (K-fold, hold-out)
Agnifili, Luca (2023)	Glaucoma	Tree-based	Classification and Regression (CART)	None	Trabulectomy failure	N/A	None
Francis, Mathew (2023)	Cornea and external disease	Tree-based/Other	Lasso linear regression and decision tree classifier	Postoperative corneal thickness measured via CORVIS-ST	Postoperative Kc mean, ectasia	Yes	Internal (hold-out)
Hayashi, Takahiko (2023)	Cornea and external disease	Neural network	CNN	None	Need for rebubbling	N/A	External
Joseph, Naomi (2023)	Cornea and external disease	Tree-based	Random Forest	Logistic regression	Corneal graft rejection	No	Internal (K-fold, holdout)
Patefield, Alastair (2023)	Cornea and external disease	Neural network	CNN	Clinician	Corneal graft detachment	Yes	Internal (K-fold)
Triepels, Ron (2023)	Cataract and anterior segment	Neural network	MLP	Logistic regression	Posterior capsule rapture	Yes	Internal (K-fold, hold-out)
Ang, Marcus (2022)	Cornea and external disease	Tree-based	Random Forest	Logistic regression	Corneal graft survival	Yes	Internal (bootstrap)
Banna, Hasan (2022)	Glaucoma	Tree-based	Random Forest	Logistic regression	Trabulectomy failure	Yes	Internal (K-fold)
Muijzer, M.B. (2022)	Cornea and external disease	Tree-based	Random Forest	None	Graft detachment	N/A	Internal (K-fold)
Patel, Sanjay (2022)	Cornea and external disease	Tree-based	Gradient boosting	None	Postoperative corneal thickness	N/A	External
Santhiago, Marcony (2022)	Cornea and external disease	Tree-based	Gradient boosting	Percent tissue altered, ectasia risk score system, residual stromal bed, corneal thickness	Ectasia	Yes	Internal (K-fold)
Sun, Yiming (2022)	Oculoplastic and orbital surgery	Neural network	Pix2Pix conditional GAN	None	Postoperative appearance after blepharoptosis surgery (overlap ratio, multiple radial midpupil lid distances)	N/A	None
Wang, Jinhong (2022)	Cataract and anterior segment	Neural network	CTT	None	Postoperative visual acuity	N/A	Internal (K-fold)
Hayashi, Takahiko (2021)	Cornea and external disease	Neural network	CNN	None	SBB formation	N/A	Internal (K-fold)
Lanza, Michele (2021)	Cataract and anterior segment	Tree-based	Decision tree	Logistic regression	Postoperative complications∗	Yes	None
Yoo, Tae Keun (2020)	Oculoplastic and orbital surgery	Neural network	Conditional GAN	None	Postoperative appearance after orbital decompression surgery	N/A	Internal (K-fold)
Gupta, Manish (2016)	Cornea and external disease	Tree-based	GDBT regression	None	Postoperative uncorrected visual acuity	N/A	Internal (K-fold)
Mohammadi, Seyed-Farzad (2012)	Cataract and anterior segment	Neural network	40 neurons	Logistic regression	Posterior capsule rapture	Yes	None

AI = artificial intelligence; ANN = artifical neural network; CNN = convolutional neural network; CORVIS-ST = corneal visualization scheimpflug technology; CPC = cyclophotocoagulation; CTT = cross-token transformer; GAN = generative adversarial network; GDBT = gradient descent boosting-based decision tree; MIGS = microinvasive glaucoma surgery; MLP = multilayer perceptron; SBB = successful big-bubble; XGBoost = extreme gradient boosting.

Figure 3.

Scatter plot of AI versus comparator performance (AUC) across subspecialties. AI = artificial intelligence; AUC = area under the curve.

Validation and Calibration

Internal validation (cross-validation or bootstrapping) was performed in 75% (n = 21) of studies. Only 3 (11%) studies incorporated external validation, which involves testing a trained model on an entirely independent data set. Calibration metrics were infrequently reported, and no study assessed practical model deployment or the impact on clinical decisions.

Performance by AI Model Type

Tree-Based Models

Random forest, gradient boosting, and Extreme Gradient Boosting models, which fall under the tree-based algorithms domain, accounted for the majority of models investigated.5, 6, 7, 8, 9^,11,15,18, 19, 20, 21^,24,25,27,28 A range of 6–432 input features was integrated to yield prognostic outputs. These features included demographic and clinical parameters (e.g., age, sex, IOP, and keratometry), and imaging-derived data such as endothelial cell density.

When assessing the 9 studies that involved comparison of tree-based algorithms to conventional forecasters such as logistic regression statistical methods and clinician judgment of postoperative recovery and risk, 8 (88.9%) reported superior predictive performance for AI. These studies achieved low-to-high predictive accuracy, with AUC values ranging from 0.55 to 0.92 (n = 7). However, conclusions drawn from this range should be interpreted with caution, given the limited number of studies.

DL Models

Following tree-based models, DL architectures were prominently featured in the included studies, notably through convolutional neural networks, multilayer perceptrons, and generative adversarial networks.12, 13, 14^{,16,17,22,23,26,}29, 30, 31, 32 These models typically operated on high-dimensional raw inputs such as images (e.g. OCT) from which relevant features were extracted implicitly without requiring manually specified input variables. Only 3 DL studies included a direct performance comparison to conventional methods. However, among these studies, AI was reported to outperform the comparators. In this domain, reported AUCs ranged from 0.63 to 0.99 (n = 7). This represents a broad range of values, likely reflecting differences in imaging inputs and outcome definitions. Moreover, similar to the reporting on tree-based model performance, considerable caution should be exercised in the interpretation of this range given the limited number of included studies.

Performance by Surgical Domain

Cornea and External Disease

Within this subspeciality domain, AI-based models were most frequently applied to predict graft survival or detachment after keratoplasty procedures such as penetrating keratoplasty, deep anterior lamellar keratoplasty, or Descemet membrane endothelial keratoplasty (n = 7; 25%).^6,15,17,19,24, 25, 26 Other studies assessed outcomes such as postoperative visual acuity, central corneal thickness, or the need for rebubbling after the procedure. In this domain, reported AUCs ranged from 0.63 to 0.99, with some of the highest-performing models (AUC >0.90) identified (Fig 2).

Glaucoma

Seven studies (25%) focused on filtration procedures, including trabeculectomy, microinvasive stent implantation, and tube shunts.^5,7, 8, 9^,12,21,22 Definitions of surgical success varied but commonly included IOP reduction and maintenance of visual acuity. Reported AUCs ranged from 0.68 to 0.93, with AI models generally outperforming comparator methods.

Cataract and Oculoplastic Surgery

Studies in these domains were fewer (n = 7; 25%) and more exploratory.^16,20,23,29, 30, 31, 32 In cataract surgery, models were used to predict postoperative visual acuity and complications, such as posterior capsule rupture or opacity. Oculoplastic studies examined anatomic and aesthetic outcomes after blepharoptosis repair or orbital decompression, demonstrating proof-of-concept feasibility. Among the included studies in this domain, Sun et al employed a DL model, which achieved a mean overlap ratio of 0.858, with 95% of predicted eyelid positions following blepharoptosis surgery falling within 1 mm of actual postoperative marginal reflex distance.³⁵ Further, a DL model was deployed by Yoo et al³⁸ to generate postoperative facial appearances after orbital decompression surgery for thyroid-associated ophthalmopathy, reporting a structural similarity index of 0.655.

Comparator Versus AI Performance Across Cornea and Glaucoma Surgical Domains

As shown in Figure 3, most studies clustered above the line of equivalence, indicating that AI models achieved comparable or superior performance (as measured by AUC) to comparator methods, encompassing traditional statistical models such as logistic regression or Cox proportional hazards, and clinician judgement. Corneal studies, such as those by Patefield et al and Santhiago et al, demonstrated the clearest performance gains. Notably, Patefield et al was the only study in this AUC comparison which employed clinician judgement as the comparator, whereby a cornea specialist undertook a “yes/no” questionnaire regarding the predicted occurrence of graft detachment based on OCT scans and limited clinical information.^32,34 Meanwhile glaucoma studies, including those by Banna et al and Lee et al,^13,27 showed more modest but consistent improvements. No studies reported the inferior performance of AI relative to conventional comparators.

Risk of Bias and Applicability

Supplementary Fig S2, available at www.ophthalmologyscience.org, presents a Prediction Model Risk of Bias Assessment Tool-AI assessment of model development for included studies across 2 dimensions: quality (Supplementary Fig S2A) and applicability (Supplementary Fig S2B). For quality of model development, 21 out of 28 studies (75%) were gauged to have a high overall risk of bias, driven by factors such as small samples, single center-derived data sets of ethnically homogenous patients, and retrospective design. In the predictors domain, many models utilized input features extracted from retrospective records (e.g. age, IOP, and smoking history) or imaging data sets (e.g. anterior-segment OCT). Moreover, missing data handling was infrequently reported. This lack of standardization and multimodality of predictor input features undermines the future reproducibility of these AI models.⁴² Further subjectivity was introduced in some studies, which warranted greater concern about bias in this domain. For example, in the study by Birla et al,¹⁵ the intraoperative predictor of Tenon’s thickness was subjectively graded by the operating surgeon without standardization. In the study by Lee et al,²⁷ the researchers categorized all psychiatric medications into a single predictor category to adjust for the unusually high prevalence of patients taking these medications in their sample, limiting application to other data sets due to introduced categorical subjectivity.

Within the included studies, surgical outcomes were defined inconsistently, introducing significant bias in the outcome domain. For example, in 2 studies which investigated the surgical failure of trabeculectomy as their predicted outcome of interest, Birla et al defined failure as postoperative IOP ≤18 mmHg or ≥50% reduction in IOP compared to baseline, while Barry et al defined failure upon either one of the following criteria being met: (1) IOP >80% of the preoperative IOP on 2 consecutive visits, (2) patients on more classes of glaucoma medications postoperatively than preoperatively, or (3) a second glaucoma surgery or revisional surgery was performed.^14,15 Moreover, some studies incorporated insufficient follow-up periods to capture clinically meaningful events. For example, in a study by Sun et al, which investigated a model to predict postoperative appearance after blepharoptosis surgery, the mean follow-up period was 115.7 days.³⁵ In comparison to other study protocols, this follow-up duration may be insufficient in accurately assessing the predictive ability of the AI model to predict the cosmetic and anatomical outcomes of surgery.⁴³

Applicability concerns throughout the first 3 domains (participants and data sources, predictors, and outcomes) were common, with 18 out of 28 studies being at high risk of bias (64%), reflecting limited generalizability.

Supplementary Fig S3, available at www.ophthalmologyscience.org, details the assessment of model evaluation across the risk of bias and applicability dimensions. For risk of bias, 21 of 28 studies (75%) were found to be at high risk of bias. These concerns were chiefly derived from the widespread lack of external validation and calibration assessments. For instance, k-fold validation on the same data set used to derive the model was employed in several studies, such as Gupta et al.¹⁷ Moreover, concerns pertaining to the reporting of model performance were also noted, such as post hoc exclusion of patients to report ensemble metrics in the study conducted by Kim et al.²⁴ Additionally, many studies involved small sample sizes in the evaluation of their algorithms, which introduces further concerns. To reduce the risk of overfitting, an appropriate participant-to-candidate input predictor feature ratio should be maintained. For example, in the study by Joseph et al,⁴⁴ 432 quantitative features were integrated in the model algorithm applied to a sample of 44 eyes, reflecting a limited sample size relative to model complexity.²¹ Applicability concerns were also widespread among the first 3 domains of evaluation, with 15 out of 28 studies determined to be high risk of bias.

Publication Bias Assessment

Given the limited completeness of outcome reporting and substantial heterogeneity in predictive metrics, and because <10 studies reported comparable effect sizes, formal statistical assessment of publication bias was not feasible for most endpoints (e.g., funnel plots or Egger test). Nevertheless, the predominance of studies reporting positive AI performance, often outperforming traditional comparators, suggests that selective publication or reporting bias cannot be ruled out.^45,46

Methodological and Temporal Trends

There was notable heterogeneity identified across predictor sets, validation strategies, and outcome definitions in included studies. When assessing predictor models, internal validation remained the predominant approach (n = 21, 75%), while only 3 of the included studies incorporated external data sets to validate their proposed AI models. Recent publications (2022–2025) tended to investigate ensemble or hybrid architectures and were more likely to report higher AUCs. These findings likely reflect advances in feature engineering and computational methods; however, the limited availability of external validation and prospective clinical evaluation should be considered. In conclusion, these trends underscore that the field of AI-based prediction in ophthalmology is still in its early translational stage.

Discussion

This systematic review synthesizes the current evidence regarding the performance of AI-based models in prognosticating outcomes and complications after anterior segment and adnexal surgeries.⁴⁷ Across 28 included studies encompassing >2.8 million eyes, AI models, notably tree-based and DL algorithms, generally demonstrated equal or superior predictive performance compared with conventional methods. In contrast, Joseph et al found that logistic regression achieved slightly higher accuracy than random forest at the final imaging time point before graft rejection (0.85 vs. 0.83, respectively).²¹ The random Forest model, however, achieved higher accuracy than logistic regression at earlier time points prior to rejection, supporting its potential to enable earlier intervention strategies to prevent graft failure and endothelial damage. Random Forest performance was also superior when delta-radiomics features were included, which capture quantitative changes from corneal endothelial cell images between 2 time points to prognosticate adverse outcomes. Studies frequently reported AUC values >0.85 for corneal transplantation and glaucoma filtration surgery, indicating the strongest predictive performance in the present findings. These findings suggest that AI-based models have the potential to improve perioperative risk assessment; however, the strength of these findings is constrained by substantial methodological weaknesses and limited generalizability.

External validation to assess clinical applicability was rarely implemented in study design (present in only 11%). Moreover, 75% of studies were deemed to have a high risk of bias. This was largely attributed to inadequate handling of missing data, small event counts relative to the number of predictors, and inconsistent reporting of performance metrics. Additionally, the calibration measures and confidence intervals for these findings were infrequently reported, constraining robust interpretation and reliability of these findings. The data sets in these studies were largely single-center, retrospectively collected, and geographically limited, which exacerbated overfitting and concerns about generalizability. These methodological issues are not unique to ophthalmic AI research but have similarly been highlighted in prior reviews of anterior segment AI applications, which emphasize the lack of standardized validation and transparent reporting across corneal and cataract models.^48,49 The limited number of studies incorporating external validation measures remains a persistent issue across all health domains, as most remain exploratory or proof-of-concept. Limitations in accessing data from multiple providers, concerns about data and security, and the lack of electronic medical records and semantic interoperability constitute barriers to researchers' access to diverse, multicenter data to verify and clinically validate model findings.⁵⁰

Our findings echo observations from prior narrative and systematic reviews. Artificial intelligence–based prognostication has been examined in the context of vitreoretinal surgery, including macular hole repair and retinal detachment surgery.^34,35 These procedures tend to have more uniform outcome measures, in contrast to surgeries of the anterior segment and adnexa, which involve more heterogeneous outcomes. The broader range of outcomes, including visual acuity, IOP, graft clarity, and complication profiles, complicates the synthesis of pooled performance. Similar to the present findings, this heterogeneity has been noted in earlier work.^51,52 Our results, corroborated by previous findings, underscore a consistent theme: despite promising performance in proof-of-concept studies, the application of AI models across ophthalmic subspecialities is limited by heterogeneous methodologies and limited generalizability, and their translation must be substantiated by standardized outcome reporting and robust external validation.

In the context of the broader movement toward precision medicine, AI offers distinct advantages by integrating multimodal predictor variables, including demographic, imaging, and surgical factors.⁵³ For instance, AI models may quantify more subjective postoperative metrics, such as eyelid symmetry after ptosis surgery, to improve the reproducibility of findings and reporting.⁵⁴ Moreover, unlike traditional statistical approaches, AI algorithms can better integrate numerous features to model complex, non-linear relationships. This is especially pertinent in predicting postoperative outcomes after glaucoma surgery, where preoperative IOP, number of medications, drainage angle anatomy, and race are some of the baseline characteristics considered in the evaluation of anticipated surgical outcomes.⁵⁵ Such capabilities could strengthen existing risk-stratification frameworks and refine perioperative counseling to be more accurate and personalized for the patient. These innovations are particularly relevant in health care systems facing resource constraints, where optimized triage, targeted follow-up, and efficient resource allocation can enhance both efficiency of the system and patient outcomes.⁵⁶

Nevertheless, several challenges remain. Model transparency and interpretability, the so-called “black box” phenomenon, continue to limit clinical adoption. Medical liability is a key area of discussion concerning the implementation of AI-based models in health care contexts, given that current knowledge is limited to understanding these algorithms through their inputs and outputs, rather than the details surrounding their internal mechanisms. This raises concern when clinicians are precluded from justifying how an AI-derived treatment recommendation was generated, while remaining accountable for their decisions.⁵⁷ Additionally, training data derived from electronic health records and image repositories may inadvertently encode and operate upon pre-existing health disparities at the systems level, posing a risk of bias concerns.⁵⁸ Thus, addressing these issues is critical to developing ethically robust and validated AI systems for accurate postoperative prediction.

Deployment in the clinical setting requires robust external validation across diverse, large data sets, as well as greater transparency about the limitations of these data sets to inform inferences drawn from the training and validation of an algorithm.^59,60 Ideally, external validation should be performed on data sets collected from different institutions, patient populations, and surgical settings than those used for the development of the model. However, the scale of data dissemination needed for the effective validation and integration of AI models into the clinical context raises concerns around patient confidentiality and cybersecurity, driving hesitation around wide-scale data sharing.⁶¹ While the size of data sets is steadily increasing in the age of digitalization of medical information, data collection is still not performed in a standardized manner. The lack of concordance in outcome definitions further impairs the use of diverse data in AI training and validation.⁶⁰ Embedding algorithms into electronic medical record systems for automated data extraction may enhance clinical workflow, but such adoption is further complicated by disparate electronic medical record systems, an additional operational hurdle restricting the interoperability of AI systems.⁶² Calibration examined across the range of predicted values and at each relevant time point, as well as discrimination metrics for models incorporating binary and time-to-event outcomes, should be consistently reported in the assessment of a model’s predictive performance and clinical utility.⁶³

Various ethical and equity factors should also be considered when discussing the path to clinical integration. Given that AI algorithms are trained on historical data, prognostic models may exacerbate inequities experienced by populations subjected to systemic bias and health disparities reflected in the training data sets. Therefore, due diligence must be taken to identify vulnerable patient populations and ensure adequate representation in numbers and characteristics. As most available data sets used to train AI models are derived from studies that have historically been biased toward participants from Western and socioeconomically advantaged backgrounds, AI models can propagate inaccurate and biased predictions when applied to diverse patient populations.⁶⁴ Continuous postdeployment monitoring is also recommended to ensure real-time feedback is collected from clinicians and patients to effectively address evolving ethical concerns and consider alternative labels.⁶⁵ Overall, the path to streamlined adoption of AI in clinical settings should prioritize showing clinical benefit rather than merely statistical significance. Such a path involves the progression from proof-of-concept studies, followed by studies employing external validation and subsequent prospective evaluation of the algorithm, to eventual commercial distribution with ongoing monitoring and evaluation.^61,66

This review has several limitations worth noting. Many of the included models were developed using small training cohorts, which increases the risk of overfitting. External validation was infrequent, further limiting generalizability and precluding definitive conclusions regarding the comparative model's performance. This marks a critical step in deploying an AI model in high-risk settings such as health care, where performance is evaluated on a data set derived from cohorts and facilities distinct from those used to train the model. Model performance on external validation data sets tends to decline relative to performance on the original training data sets.⁶⁷ In a phenomenon known as optimism bias or overfitting, AI models are optimized to the specific characteristics of their training data, encountering unfamiliar variation in external data, which can translate to a decay in performance.⁶⁸ Thus, it is essential to discern whether a model can perform consistently across diverse clinical settings and meets acceptable patient-safety standards. Moreover, heterogeneity in outcome definitions and sparsity of reporting calibration statistics precluded conducting a meta-analysis in the present study. Given this, the reported AUCs and their interpretations should be considered exploratory, emphasizing trends rather than definitive estimates. Lastly, publication bias cannot be excluded given the predominance of studies reporting favorable AI performance.

Large, multicenter prospective studies employing external validation assessments are necessary to definitively establish the clinical utility of AI prognostic models in ophthalmic surgery. Future research should aim to report findings in a standardized manner to enable meaningful comparison and synthesis of AI model performance. A consistent set of reporting metrics should be established, including discrimination measures (e.g. AUC, C-statistic), calibration measures (e.g. calibration slope, calibration plot), and where applicable, accuracy, sensitivity, specificity, and positive and negative predictive values. Moreover, outcomes should be defined using validated, clinically meaningful endpoints with specified measurement time points. Further, adherence to established reporting frameworks such as the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis – AI would substantially maintain the consistency, transparency, and comparability of future studies, and the inclusion of health economics analyses would further guide policy and resource allocation regarding AI integration into clinical workflows.^40,42 Collaborative efforts to establish open, diverse, and well-annotated data sets will also accelerate reproducibility and clinical translation.⁶⁹

Conclusions

Artificial intelligence models may have a role in predicting postoperative outcomes after anterior segment and ocular adnexal surgeries, often outperforming traditional statistical methods. However, these algorithms should be regarded as adjunctive tools rather than replacements for expert clinical judgment, particularly given that the current evidence base is limited by small sample size, inadequate external validation, and inconsistent reporting standards. Continued collaboration between data scientists and ophthalmic surgeons is essential to develop transparent, clinically appropriate models that can be safely integrated into clinical practice and have an impact on perioperative management. Moreover, future research in this domain would perhaps benefit from a mixed-methods approach that integrates quantitative and qualitative evidence to provide a more comprehensive understanding of AI model performance and clinical translatability. Ultimately, the clinical transition of AI-driven prognostication will depend on rigorous validation efforts across diverse populations to translate to tangible benefits for patients and the wider health system.

Manuscript no. XOPS-D-26-00431.