Keratoconus is a bilateral, usually asymmetric corneal ectasia characterized by progressive stromal thinning, corneal steepening, irregular astigmatism, and reduced visual quality. It commonly becomes clinically apparent during adolescence or early adulthood, although progression can occur across a broader age range. Population estimates vary markedly according to case definition and ascertainment method; established risk factors include eye rubbing, family history, allergy, asthma, and eczema [1-3].

Contemporary KC assessment relies on clinical examination together with corneal topography and tomography, with AS-OCT and biomechanical assessment increasingly contributing to the evaluation of early or ambiguous cases. The main clinical challenge is not only recognition of manifest disease but also discrimination of subclinical or forme fruste disease and timely identification of progression before substantial visual deterioration occurs [1,2].

AI offers a framework for integrating large numbers of

corneal variables and image-derived features that may be difficult to interpret simultaneously by conventional rule-based systems. Machine-learning approaches include logistic models, decision trees, support vector machines, random forests and gradient-boosting methods, while deep-learning approaches use multilayer neural networks and convolutional neural networks to learn image features directly. Recent reviews show consistently high performance for manifest KC, but they also highlight weaker evidence for early disease, heterogeneous labels and reference standards, and incomplete validation outside the development setting [4-8].

This mini-review therefore focuses on AI-supported diagnosis, early detection, monitoring, and progression prediction in keratoconus, with emphasis on clinically relevant performance, imaging inputs, current limitations, and priorities for translation into routine practice.

A focused narrative search was undertaken for English-language literature published between January 2020 and December 2025. PubMed/MEDLINE and the Cochrane Library were searched using combinations of the terms “keratoconus,” “artificial intelligence,” “machine learning,” “deep learning,” “corneal topography,” “corneal tomography,” “anterior segment optical coherence tomography,” “biomechanics,” and “progression.” Reference lists of recent systematic reviews and meta-analyses were additionally screened. Priority was given to systematic reviews, meta-analyses, large diagnostic studies, and studies evaluating progression or clinically relevant multimodal models. Evidence was synthesized narratively because of substantial heterogeneity in devices, case definitions, AI architectures, reference standards, and validation strategies [4-8,12]. AI in Keratoconus Diagnosis Diagnostic Performance: Meta-Analytic Evidence The strongest synthesis of diagnostic accuracy is the Cochrane diagnostic test accuracy review by Vandevenne et al. [4]. It included 63 studies published through 2022 and, importantly, reported three separate units of analysis: 44 studies analyzed 23,771 eyes, four studies analyzed 3,843 participants, and 15 studies analyzed 38,832 images. This distinction corrects the misleading interpretation that these quantities represented a single nested sample. For manifest KC, AI showed a summary sensitivity of 98.6% (95% CI 97.6%-99.1%) and specificity of 98.3% (95% CI 97.4%-98.9%). For subclinical KC, summary sensitivity was lower at 90.0% (95% CI 84.5%-93.8%) while specificity was 95.5% (95% CI 91.9%-97.5%). In analyses that combined manifest and subclinical disease, reported sensitivity and specificity were 96.2% and 98.0%, respectively. However, the certainty of evidence was low to very low because most studies had important risks of bias, especially from case-control sampling, inconsistent reference standards, and limited real-world validation [4]. Independent meta-analyses support the same overall pattern. Cao et al. [5] found pooled sensitivity and specificity of 97.0% and 98.5% for established KC, but 88.2% and 94.7% for early KC. Bodmer et al. [7] similarly reported pooled sensitivity of 97.5% and specificity of 97.2% for deep-learning models using topographic images, while emphasizing modest methodological quality. These findings indicate that excellent performance for clearly established disease should not be extrapolated directly to screening populations or subtle preclinical disease. Figure 1. Summary sensitivity and specificity reported for AI-assisted keratoconus detection. Values are derived from the Cochrane review and its reported mixed-disease analysis [4]. Imaging Modalities and Algorithmic Approaches AI models have been developed using Placido-disc topography, Scheimpflug tomography, AS-OCT, epithelial and stromal thickness maps, and biomechanical data. A systematic review comparing imaging modalities found that combined imaging approaches can improve discrimination, particularly for early disease; however, performance estimates vary according to disease definition, dataset composition and validation design [6]. Tomography remains one of the most frequently used inputs because it provides anterior and posterior curvature, elevation, pachymetry and spatial thickness-distribution information. Deep-learning systems can process raw or color-coded tomographic maps, whereas conventional machine-learning models commonly rely on device-generated indices. AS-OCT adds high-resolution structural information, including epithelial and stromal thickness patterns that may partially mask or reveal early ectatic change [6,8,9]. Biomechanical variables may add complementary information. Quanchareonsap et al. [10] evaluated 1,668 tomographic images and 611 biomechanical images and reported an AUC of 0.938 for a Pentacam-map model. Adding Corvis ST dynamic corneal response information increased AUC to 0.985, and incorporation of the corneal biomechanical index produced an AUC of 0.991. These results support multimodal integration, although prospective external validation remains necessary before such models can be generalized across populations and devices. Figure 2. AUC of sequential AI models using tomographic and biomechanical inputs in the diagnostic study by Quanchareonsap et al. [10]. DCR, dynamic corneal response; VSP, Vinciguerra screening parameters. Early and Subclinical Detection Early KC remains the most clinically demanding target for AI because structural abnormalities are smaller, diagnostic labels are less consistent, and control groups may include eyes with occult susceptibility. The lower pooled sensitivity for early disease in both the Cochrane review and the meta-analysis by Cao et al. underscores this limitation [4,5]. Definitions such as subclinical KC, forme fruste KC, keratoconus suspect, and very asymmetric ectasia should therefore be interpreted according to the diagnostic criteria used in each study rather than treated as completely interchangeable categories. Large contemporary datasets nevertheless demonstrate that AI can assist early detection. Alió Del Barrio et al. [9] developed an artificial neural network using 6,677 eyes assessed with a combined Placido-disc and AS-OCT topographer. The network achieved 98.6% accuracy for established KC. For keratoconus-suspect eyes, overall accuracy remained 98.5%, but recall was 69.7%, demonstrating that high overall accuracy can coexist with clinically important false-negative risk in the most difficult subgroup. This distinction is especially relevant for refractive-surgery screening, where sensitivity to subtle ectasia risk is critical. Clinical Decision Support The near-term clinical role of AI is best framed as decision support. Algorithms can standardize image interpretation, highlight atypical patterns, provide probability-based outputs, and support triage in settings where corneal subspecialty expertise is limited. The 2024 systematic review by Goodman and Zhu [8] identified 93 original studies, including 61 focused on KC or subclinical KC detection, 16 on severity or clinical-feature assessment, seven on progression prediction, and six on treatment-response characterization. This breadth shows that AI is expanding beyond binary diagnosis toward staging, monitoring, and management support. For clinical use, however, model output should be interpreted alongside examination findings, refraction, serial imaging, age, risk factors, and treatment context. A model with excellent discrimination can still be unsafe if it is poorly calibrated, trained on a narrow population, or applied to data from a different device. Clinical decision support therefore requires transparent reporting of the intended population, reference standard, threshold selection, calibration, and validation setting [4,7,8]. AI in Keratoconus Monitoring and Progression Prediction The Clinical Challenge of Progression Progression is clinically more difficult to predict than established disease is to diagnose. Serial assessment typically considers changes in anterior or posterior curvature, Kmax, pachymetry, and multivariable progression displays. Measurement noise, age-related differences in progression risk, contact-lens effects, and inconsistent definitions of clinically meaningful change complicate model development and comparison [1,2,12]. Evidence for AI-Based Progression Prediction Kamiya et al. [11] evaluated deep learning using six color-coded AS-OCT maps in 218 keratoconic eyes. The combined-map approach achieved an accuracy of 0.794 for distinguishing progressive from non-progressive KC. Posterior elevation produced the highest single-map accuracy (0.798), and adjustment by age subgroup increased overall accuracy to 0.849. The study demonstrates the value of combining spatial imaging patterns with a clinically meaningful variable such as age. A 2025 systematic review by Hashemi et al. [12] synthesized ten studies involving 10,940 eyes and found AUC values ranging from 0.77 to 0.93 and accuracies from 77.5% to 84.9%. Posterior elevation, Kmax and younger age were repeatedly identified as important predictors. Some models also incorporated non-imaging variables such as eye rubbing and selected laboratory or nutritional factors; these additional predictors require further validation before routine clinical use. Importantly, the review found that the included studies relied predominantly on internal splitting or cross-validation rather than independent multicenter external validation [12]. Taken together, progression models appear promising but less mature than diagnostic models. Their most defensible current application is risk stratification to guide follow-up intensity and support, rather than replace, serial clinical assessment and established progression criteria. Multimodal Risk Stratification and Treatment Decisions The direction of development is toward multimodal models that combine raw imaging, device-derived indices, refraction, age and other clinical variables. Such models are attractive because KC progression is a biological and geometric process that is unlikely to be captured optimally by one corneal parameter. Multimodal integration may also reduce overdependence on Kmax, which does not fully reflect posterior surface change or localized thinning [8,10-12]. For corneal cross-linking decisions, AI should be used cautiously. A prediction model may identify a patient as high risk, but treatment decisions require consideration of documented change, age, visual function, corneal thickness, measurement repeatability, patient preferences, and the potential consequences of delayed treatment. The appropriate target is therefore clinically interpretable risk support rather than automatic treatment authorization. Post-Refractive Ectasia Risk AI-based ectasia screening has direct relevance to refractive surgery because subtle preoperative abnormalities can be difficult to recognize with single indices. Models that integrate anterior and posterior corneal shape, pachymetric distribution, epithelial information, and biomechanics may improve identification of eyes at risk. Nevertheless, the evidence from KC diagnostic studies should not be assumed to provide equivalent accuracy for predicting post-refractive ectasia, which is a distinct clinical endpoint requiring appropriately designed longitudinal validation [6-8]. Key Predictive Factors and Biomarkers Across diagnostic and progression studies, several recurring variables contribute to AI classification. Their importance depends on imaging platform, disease definition, feature engineering and model architecture. • Tomographic and topographic features: Kmax, anterior and posterior curvature, elevation parameters, asymmetry indices, and spatial pachymetric progression. • Posterior corneal features: posterior elevation is repeatedly associated with early ectatic change and has shown strong contribution in progression models [11,12]. • Thickness-related features: thinnest pachymetry, regional thickness distribution, and epithelial or stromal thickness patterns obtained with AS-OCT [6,8,9]. • Biomechanical features: dynamic corneal response variables and composite indices from air-puff deformation analysis can add complementary information to tomography [10]. • Clinical variables: younger age is repeatedly associated with progression risk, while eye-rubbing and atopic history are clinically relevant risk factors that may improve contextual interpretation [2,3,12]. The strength of evidence is greatest for corneal geometry, thickness, biomechanics and age. Other proposed biomarkers should be interpreted cautiously until their incremental predictive value is replicated prospectively in independent cohorts. Table 1. Selected contemporary evidence on AI for keratoconus diagnosis and progression Study Design / dataset AI application Key finding Cao et al., 2022 [5] Systematic review/meta-analysis; 35 studies ML for established and early KC KC: sensitivity 97.0%, specificity 98.5%; early KC: sensitivity 88.2%, specificity 94.7% Vandevenne et al., 2023 [4] Cochrane DTA review; 63 studies Topography/tomography/OCT AI diagnosis Manifest KC: 98.6% sensitivity, 98.3% specificity; subclinical KC: 90.0% sensitivity, 95.5% specificity Hashemi et al., 2024 [6] Systematic review/meta-analysis; 36 studies Comparison of corneal imaging modalities High pooled accuracy across Scheimpflug, Placido and combined imaging; performance differed by disease stage and modality Bodmer et al., 2024 [7] Systematic review/meta-analysis; 19 studies Deep learning in diagnostic work-up Topography-input pooled sensitivity 97.5%, specificity 97.2%; methodological quality limited Goodman & Zhu, 2024 [8] Systematic review; 93 original studies Diagnosis, severity, progression and treatment response 61 diagnostic studies; 7 progression studies; 6 treatment-response studies Alió Del Barrio et al., 2024 [9] Multicenter cross-sectional; 6,677 eyes Placido + AS-OCT ANN KC accuracy 98.6%; suspect-eye recall 69.7% Quanchareonsap et al., 2024 [10] Diagnostic study; 1,668 tomographic and 611 biomechanical images Pentacam ± Corvis ST deep learning AUC increased from 0.938 (tomography) to 0.991 with added biomechanical index Kamiya et al., 2021 [11] Progression study; 218 eyes Deep learning of six AS-OCT maps Overall accuracy 0.794; age-adjusted accuracy 0.849 Hashemi et al., 2025 [12] Systematic review; 10 studies, 10,940 eyes AI prediction of progression AUC 0.77-0.93; accuracy 77.5%-84.9%; external validation insufficient CHALLENGES AND LIMITATIONS Heterogeneous Disease Definitions and Reference Standards A central limitation is the lack of uniform definitions for early KC and progression. Studies differ in whether labels are based on clinical signs, device classifications, expert consensus, fellow-eye status, change in Kmax, pachymetric change, or combinations of parameters. When the reference standard itself is variable, apparently precise AI performance estimates may not be directly comparable [4-8,12]. Patient Selection, Spectrum Bias and Class Imbalance Many diagnostic studies compare clearly abnormal KC eyes with clearly normal controls. This two-gate case-control design exaggerates separation relative to real clinics, where patients often have borderline findings, prior surgery, contact-lens effects or coexisting corneal disease. Enriched datasets and class imbalance can also make accuracy appear high even when sensitivity in the clinically important minority class is inadequate [4,5,7,9]. Limited External and Prospective Validation Internal train-test splits and cross-validation are useful during model development but do not establish transportability. Diagnostic and progression reviews consistently identify limited independent external validation, and the 2025 progression review found no independent multicenter external validation among its included studies [7,8,12]. External testing should include different institutions, ethnic groups, disease prevalence, operators and device generations. Device Dependence and Data Shift Most models are developed using data from specific topographers, tomographers or OCT systems. Differences in acquisition geometry, preprocessing, software versions, normative databases and map rendering can cause data shift and performance deterioration. Platform-independent models, harmonized inputs and explicit cross-device validation are therefore important priorities [6-8,10]. Calibration, Explainability and Clinical Utility AUC, sensitivity and specificity describe discrimination but do not fully establish clinical usefulness. Probability calibration, predictive values at realistic disease prevalence, threshold selection, net benefit and clinical-impact analysis are necessary when outputs are used to determine referral intensity or intervention. Explainable AI methods may help clinicians understand which corneal regions or parameters drive a prediction, but interpretability should complement rather than substitute for robust validation. Reporting, Reproducibility and Governance Reproducibility is limited when datasets, preprocessing procedures, code, threshold selection or model architectures are incompletely described. Future work should follow contemporary reporting standards for prediction and diagnostic AI studies, document patient flow and exclusions, prevent leakage between training and test sets, and address privacy, cybersecurity and regulatory oversight before deployment in routine care. Future Directions Future studies should move from proof-of-concept classification toward prospective clinical validation. Priority areas include multicenter recruitment; consistent definitions of manifest, subclinical and progressive KC; prespecified external test cohorts; calibration and decision-curve analysis; and head-to-head comparison with experienced clinicians and established device indices. Multimodal models combining tomography, AS-OCT, epithelial mapping, biomechanics and clinical risk factors are particularly promising because they reflect complementary aspects of ectatic disease. However, greater model complexity is justified only when it produces clinically meaningful incremental benefit and remains interpretable, reproducible and feasible across devices and settings [8,10-12]. Integration into clinical workflows should emphasize transparent risk estimates, quality-control warnings, automated comparison with prior scans, and clinician oversight. In underserved settings, validated AI could support referral and triage where subspecialty expertise is limited, but deployment should be accompanied by local validation and clear pathways for confirmatory specialist assessment.

Artificial intelligence has demonstrated high diagnostic discrimination for manifest keratoconus and useful, although less consistent, performance for early disease and progression prediction. The best-supported evidence shows summary sensitivity and specificity above 98% for manifest KC, while early/subclinical disease remains more difficult and progression models generally show moderate-to-high rather than near-perfect discrimination [4,5,12].

The most promising direction is multimodal integration of corneal shape, thickness, OCT and biomechanical information with relevant clinical variables. Nevertheless, current evidence is constrained by heterogeneous definitions, case-control sampling, device dependence, incomplete reporting and limited external validation. AI should therefore be viewed as a clinical adjunct that can standardize interpretation and support risk stratification, not as a replacement for comprehensive corneal assessment or documented longitudinal change. High-quality prospective multicenter validation will determine whether these models can safely improve early detection, monitoring and treatment decisions in routine keratoconus care.