Keratoconus is a progressive, non-inflammatory ectatic disorder of the cornea characterized by stromal thinning, anterior protrusion, and irregular astigmatism, collectively resulting in significant visual impairment.⁽¹⁾ Accurate and early identification of keratoconus, including its preclinical forms and susceptibility in refractive surgery candidates, remains a central clinical challenge. Over the past three decades, the field has been profoundly shaped by the pioneering contributions of Dr. Yaron Rabinowitz, whose systematic, quantitative approach to corneal topography analysis established a scientific framework that continues to support contemporary ectasia screening.^(2,3)

Among Dr. Rabinowitz's most enduring contributions are the KISA% index ⁽²⁾, which integrates keratometry (K), I-S value, skewed radial axis (SRAX), and astigmatism into a single diagnostic score; the inferior-superior (I-S) value, which quantifies the dioptric asymmetry between inferior and superior corneal zones;⁽³⁾ and the pachymetry-adjusted I-S (PA/I-S) index, which normalizes the I-S value for corneal thickness, thereby adding a structural dimension to functional asymmetry assessment.⁽⁴⁾ These indices exemplify his vision of transforming complex clinical data into actionable, reproducible diagnostic criteria.

The emergence of Artificial Intelligence (AI) offers a complementary path for extending the utility of these established indices.⁽⁵⁾ The Applied Artificial Intelligence methodology (AAI or A²I), as defined in this work, is built upon three core pillars: foundation (grounding in domain-specific scientific knowledge), methodology (rigorous application of Knowledge Discovery in Databases, KDD), and competence (mastery of machine learning techniques and their clinical validation).^(6,7) Rather than replacing prior knowledge, this paradigm seeks to amplify it.

Building on this foundation, the present study was designed as a tribute to Dr. Rabinowitz by enhancing the PA/I-S index with AI, demonstrating how he could have applied the artificial intelligence (A²I) methodology to improve it. To accomplish this, we developed an AI predictive model called AI Tribute, using the constitutive elements of the PA/I-S index – minimum pachymetry and I-S value – as inputs, trained on normal and VAE (videokeratography-asymmetric ectasia [VAE-E] + videokeratography-asymmetric ectasia with normal topography [VAE-NT]) eyes to capture the full clinical spectrum of ectasia susceptibility. This model was then evaluated in two complementary clinical scenarios – normal versus clinical ectasia and normal versus VAE-NT – and compared to Dr. Rabinowitz's established indices, as detailed below.

The present study was designed as a scientific tribute to Dr. Yaron Rabinowitz, one of the most influential scientists in the history of keratoconus research. Instead of merely citing his work, we aimed to honor it by demonstrating, through rigorous application of the A²I methodology, how his foundational ideas can be expanded and improved with machine learning while maintaining the scientific logic that originally motivated them.^(2-4) The findings support this idea: AI Tribute, trained solely on Dr. Rabinowitz's core variables, achieved statistically significant improvements over all reference indices in both clinical evaluation scenarios – Normal vs. Clinical Ectasia and Normal vs. VAE-NT – showing the A²I approach's ability to extract latent discriminative information from established clinical knowledge.^(5,9)

Feature selection was intentionally limited to Dr. Rabinowitz's conceptual domain – topographic asymmetry and minimum pachymetry – without adding biomechanical, elevation, or other tomographic data. This restriction was not a methodological flaw but a design choice, aligned with the A²I paradigm, where domain principles direct computational exploration,⁽⁶⁾ and was driven by a specific question: how much discriminative power can be obtained from Dr. Rabinowitz's own variables when modeled nonlinearly by machine learning algorithms?

In the Normal versus Clinical Ectasia scenario, all reference indices achieved high discriminative performance, reflecting the pronounced topographic and pachymetric differences that characterize manifest ectasia. Among them, PA/I-S index (AUC = 0.978) outperformed the I-S value (AUC = 0.968), KISA% (AUC = 0.953), and pachy min (AUC = 0.938), with all pairwise differences reaching statistical significance (p < 0.05). This ordering is clinically logical: the PA/I-S formulation, by normalizing topographic asymmetry for corneal thickness, provides a more comprehensive structural representation than the I-S value alone. Its advantage over KISA% suggests that adding SRAX and astigmatism offers limited additional discriminative value in a population where ectasia is already clinically evident.

AI Tribute achieved an AUC of 0.985 in the normal vs. clinical ectasia scenario, demonstrating statistically significant improvements over all reference indices, including a gain over the already high-performing PA/I-S index. Importantly, this improvement was not achieved by adding new clinical variables: the model was trained exclusively on pachy min and I-S value – the two core elements of the PA/I-S index – and its superiority over the PA/I-S index stems entirely from the ability of the MLP architecture to model nonlinear interactions between them. While the PA/I-S formula applies a fixed linear correction for corneal thickness, the neural network learns a flexible, data-driven mapping that adjusts the importance of each input based on the local data structure, extracting discriminative information that the original formula could not capture. This aligns with broader evidence that machine learning models consistently outperform hand-crafted linear indices for ectasia detection, even when limited to the same input features.^(5,6,9) The final architecture – a single hidden layer with only three neurons – achieved near-ceiling discrimination with a simple structure, showing that complex design is not necessary for clinical usefulness, and that careful feature selection has a bigger impact than complex models.

The normal versus VAE-NT scenario shows a fundamentally different clinical challenge. VAE-NT eyes are, by definition, topographically normal or apparently normal – they do not show obvious anterior curvature changes – and their classification as ectatic depends on the bilateral asymmetry of their fellow eye, rather than on direct topographic findings of their own. In this context, all reference indices showed substantially reduced discriminative performance compared with the clinical ectasia scenario, and their relative ordering shifted in a clinically instructive way.

Pachy min (AUC = 0.762) emerged as the top-performing reference index, surpassing the I-S value (AUC = 0.730). Meanwhile, the PA/I-S index (AUC = 0.688) scored below both, and KISA% (AUC = 0.633) showed the worst performance. The reversal of the relationship between PA/I-S and I-S is particularly revealing: although the pachymetric correction in the PA/I-S formula enhances sensitivity for manifest ectasia by penalizing abnormally thin corneas, it paradoxically decreases sensitivity in VAE-NT eyes, where corneal thickness may still be normal. In such cases, the correction reduces rather than accentuates the topographic signal. Conversely, the advantage of pachy min over the I-S value in the VAE-NT group – although not statistically significant (p = 0.0609) – reflects established evidence that subclinical structural compromise, even at the earliest stages, tends to involve pachymetric redistribution before overt curvature changes appear.^(5,8,9) KISA%, which combines multiple topographic parameters including SRAX and astigmatism, performed the worst, suggesting that in the absence of clear anterior surface abnormalities, these composite scores increase noise rather than signal.

In the normal versus VAE-NT scenario, AI Tribute achieved an AUC of 0.791, maintaining statistically significant superiority over all reference indices, with gains over pachy min, I-S value, PA/I-S index, and KISA%. The performance difference between the two scenarios – AUC 0.985 versus 0.791 – is not due to a flaw in the model but reflects the inherent biological and epidemiological challenge of the task: detecting ectasia susceptibility in eyes that show no obvious topographic signs of disease is naturally more difficult than distinguishing normal corneas from those with clinically evident ectasia. Any index based solely on topographic and pachymetric parameters will face this limit. Notably, AI Tribute retained its relative advantage over the reference indices even in this more challenging context. This is due to the training approach: by exposing the model to the full VAE spectrum – including VAE-NT eyes in the positive class – the KDD framework ensured that the learned decision boundary was optimized for the subtle patterns associated with early ectasia susceptibility, rather than only for the pronounced deformations typical of active disease. Therefore, the model captured the clinical insight that ectasia exists on a spectrum and that the boundary between health and susceptibility provides more meaningful training targets than the boundary between health and established disease.

This study has certain limitations. Both the training and evaluation were conducted using the same institutional dataset; targeted external validation in independent, multicenter cohorts is essential to verify the model's generalizability across populations with different demographic profiles and disease prevalence. Additionally, the VAE-NT classification is a clinical construct based on bilateral topographic asymmetry of the fellow eye rather than on direct longitudinal confirmation of ectasia progression; prospective studies documenting the natural history of VAE-NT eyes would strengthen the validity of this label as a training and evaluation target. The training strategy – normal versus VAE (VAE-E + VAE-NT) – was intentionally designed to maximize sensitivity at the clinically relevant decision boundary between health and susceptibility to ectasia, including its topographically silent form; the two evaluation scenarios offer complementary insights into the model's performance across the ectasia severity spectrum. Future research should focus on incorporating these models with biomechanical indices and the biomechanical impact of laser vision correction surgery on the cornea,⁽¹¹⁾ which provides a complementary perspective on ectasia susceptibility not captured by topographic or pachymetric variables alone, and integrating them into clinical decision-support tools to promote practical application.

The integration of the A²I methodology with Dr. Yaron Rabinowitz's scientific contributions showed that his fundamental indices for keratoconus detection can be improved through machine learning modeling. AI Tribute, trained exclusively on the constitutive elements of the pachymetry-adjusted/inferior-superior index – minimum pachymetry and inferior-superior value – and evaluated in two clinically distinct scenarios, achieved statistically superior discriminative performance over all reference indices in both comparisons. In the Normal versus Clinical Ectasia scenario, it reached an AUC of 0.985 with high sensitivity (94.32%) and specificity (97.74%) at the optimal cutoff. In the more challenging Normal versus videokeratography-asymmetric ectasia with normal topography scenario – where the model was applied to topographically normal fellow eyes of patients with ectatic disease – it achieved an AUC of 0.791, maintaining statistically significant superiority over the inferior-superior value, the pachymetry-adjusted/inferior-superior index, KISA%, and pachy min. These results highlight the potential of Applied Artificial Intelligence to enhance rather than replace established clinical knowledge and to respect the importance of Dr. Rabinowitz's contributions to the field of corneal ectasia screening.