Predicting the Future of Corneal Grafts: Machine Learning in Long-Term Keratoplasty Outcomes

Reframing Corneal Transplantation in the Era of Machine Learning

Corneal transplantation has become the most common form of tissue transplantation worldwide, with endothelial dysfunction as its principal indication. For decades, penetrating keratoplasty (PK), which involves replacing the full thickness of the cornea, served as the standard of care. The advent of endothelial keratoplasty (EK) procedures, particularly Descemet stripping automated endothelial keratoplasty (DSAEK), transformed the field with its promise of faster recovery, fewer complications, and potentially longer graft survival. Yet despite its rapid global adoption, questions linger regarding how well DSAEK performs compared to PK over the course of a decade, especially in populations where risk factors such as bullous keratopathy (BK) predominate.

Traditional methods of survival analysis, such as Kaplan–Meier curves and Cox proportional hazards modeling, have provided insight but are limited by their capacity to handle large, multidimensional datasets. Clinical registries spanning thousands of transplants accumulate enormous quantities of patient, donor, and procedural variables, with complex interdependencies that cannot be distilled by linear assumptions alone. This is where machine learning, and more specifically random survival forests (RSF), provide a methodological breakthrough. By leveraging decision-tree ensembles, RSF captures nonlinear interactions among dozens of variables while still generating predictive models that surpass classical regression in accuracy.

The Singapore Corneal Transplant Registry provided an ideal dataset to test these methods, covering over a decade of transplants in Asian eyes with Fuchs’ endothelial dystrophy (FED) and BK. Unlike national registries that amalgamate outcomes across varied institutions, this registry benefits from standardized surgical protocols and careful follow-up across all cases. By applying RSF to this prospective dataset, researchers were able to interrogate which donor, recipient, and procedural variables most strongly predicted long-term graft survival, setting a benchmark for integrating computational tools into ophthalmic epidemiology.

The motivation extends beyond statistics: for clinicians and patients alike, understanding which procedure offers a better long-term prognosis has direct implications for surgical planning, donor tissue allocation, and counseling patients about realistic outcomes. Machine learning is not simply a computational experiment here; it is a tool for refining the future practice of corneal transplantation in an era of high surgical demand and constrained tissue supply.

The Dataset: Ten Years of Keratoplasty in Asian Eyes

The Singapore cohort encompassed 1,335 consecutive cases, with nearly three-quarters undergoing DSAEK and the remainder PK. The population skewed toward older adults, with an average age near seventy, and included a higher burden of BK compared to FED. This distinction matters: BK, often a sequela of cataract surgery or intraocular procedures, tends to produce poorer graft survival relative to FED, which is a more isolated degenerative disorder of the endothelium. By parsing outcomes separately, the study avoided conflating these distinct risk environments.

Each case contributed dozens of variables, from donor endothelial cell density to intraoperative complications, postoperative regimen adherence, and demographic markers such as gender. This complexity is precisely where traditional models falter, as they are forced to reduce the system into linear predictors. RSF, by contrast, thrives in environments where variables may interact in subtle, nonlinear ways, such as how preoperative visual acuity modulates graft outcomes in the presence of concurrent glaucoma or systemic comorbidities. The inclusion of over forty-nine potential predictors allowed the algorithm to rank their relative contributions by variable importance scores, an output that translates complexity into actionable hierarchy.

Surgical technique itself emerged as one of the strongest determinants of survival. Ten-year outcomes demonstrated that DSAEK provided substantially higher survival than PK, with reduced rates of immune-mediated rejection, wound dehiscence, and epithelial complications. The RSF approach confirmed procedure type as one of the top-ranked variables, reinforcing what clinical intuition and shorter-term studies had long suggested. Importantly, this computational confirmation provides statistical robustness to observations that previously rested on less comprehensive analyses.

The strength of the dataset lay in its prospective nature. Unlike retrospective reviews, where case selection bias or incomplete data introduce confounders, this registry tracked outcomes in real time, recording complications, visual outcomes, and cellular metrics under uniform definitions. Such consistency enhances the credibility of machine learning results, as the model is only as reliable as the fidelity of the underlying data. In essence, the registry offered a rare convergence of high-quality longitudinal data with cutting-edge analytic methods.

Random Survival Forests as a Predictive Lens

RSF models operate by building thousands of survival trees, each using random subsets of patients and variables, then averaging their outputs to yield a consensus prediction. This ensemble method allows for robust modeling of time-to-event data such as graft failure, while reducing the overfitting that plagues single-tree methods. Variable importance plots from this study revealed that diagnosis (BK versus FED), surgical technique (PK versus DSAEK), recipient gender, preoperative visual acuity, and donor endothelial cell count were the top five determinants of graft survival. The resulting model achieved an out-of-bag C-index superior to that of traditional regression approaches, a measure of its predictive discrimination.

Beyond ranking variables, RSF also captured nonlinear associations. For example, donor endothelial cell counts exhibited a nonlinear relationship with survival probability—an effect that would be missed if constrained to linear assumptions. This observation is clinically relevant: while very low cell counts predict failure, the survival benefit plateaus beyond a certain threshold, suggesting diminishing returns in selecting only the highest-density donor tissues. Such nuances are critical for eye banks in optimizing tissue allocation without unnecessarily discarding viable grafts.

Another key insight was the gender effect. Male recipients consistently demonstrated higher rates of long-term graft failure, independent of surgical technique. While gender matching between donor and recipient has historically been debated, RSF reinforced the biological plausibility of gender-related immunologic responses influencing graft outcomes. These findings underscore the need for further immunogenetic investigations, as sex-linked differences in alloimmune reactivity may contribute to the observed disparities.

Equally important was the identification of preoperative visual acuity as an independent predictor. Poor baseline acuity often reflects advanced corneal decompensation or concomitant ocular disease. By quantifying its role as a predictor, the RSF model allows clinicians to incorporate a patient’s functional status into survival estimates. This shifts graft counseling from a one-size-fits-all prognosis to a more tailored discussion of individualized risks, a hallmark of precision medicine.

Complications and Clinical Implications

The superiority of DSAEK was not merely statistical but grounded in tangible clinical endpoints. Immune-mediated rejection, epitheliopathy, and wound instability were markedly more frequent in PK than in DSAEK. The latter, by avoiding a full-thickness wound and extensive suturing, reduced the avenues for post-operative immune activation and structural compromise. By contrast, PK patients bore the added risks of suture-related astigmatism, wound dehiscence, and greater endothelial cell loss, complications that compound over a decade to erode graft survival.

Yet DSAEK introduced its own unique challenges. Graft detachment, though less catastrophic than PK-related wound rupture, emerged as a complication intrinsic to lamellar procedures. Management of detachment requires re-bubbling procedures, which while effective, add to patient burden and clinical workload. RSF analysis contextualized these risks by demonstrating that despite such detachment events, long-term survival still favored DSAEK, affirming that its complications were generally more manageable than those of PK.

Endothelial cell attrition, a universal feature of corneal grafts, followed divergent trajectories between the two procedures. PK demonstrated a sharper decline in cell density, with implications for earlier onset of graft decompensation. DSAEK, by transplanting only the posterior lamella, preserved more donor cells and limited host inflammatory exposure. This mechanistic distinction aligns with the RSF findings, which ranked donor cell count as a high-importance variable, but one whose effect was more favorable under the lamellar surgical paradigm.

The clinical message resonates clearly: in an era where patient longevity continues to rise, the durability of a corneal graft becomes as critical as its immediate visual recovery. By validating the superiority of DSAEK over a decade, the study provides surgeons with a robust evidence base for recommending lamellar over penetrating procedures in most endothelial disease scenarios. For healthcare systems, the implication is equally strong—by investing in DSAEK infrastructure, long-term burden of repeat transplantation and complication management can be reduced.

Beyond Statistics: Toward Predictive Ophthalmology

The integration of RSF into clinical ophthalmology signals a shift from descriptive to predictive analytics. Traditional survival curves offered retrospective explanations of outcomes; RSF allows prospective estimation of risks tailored to an individual’s profile. By combining the explanatory power of Cox regression with the predictive granularity of machine learning, the study created a hybrid analytic framework that is both interpretable and actionable.

This approach also highlights the potential for expanding machine learning into other domains of transplantation. Liver, kidney, and heart graft survival each depend on a web of donor and recipient factors, where nonlinearities abound. The success of RSF in corneal transplantation may encourage analogous applications across organ systems, particularly where registry data is abundant but underutilized due to statistical complexity. Ophthalmology thus provides a proving ground for methodologies with far broader implications.

Moreover, the use of registry-based RSF analysis democratizes innovation. Randomized controlled trials remain the gold standard but are often impractical for surgical comparisons requiring decades of follow-up. By maximizing the value of prospectively collected audit data, machine learning offers a pathway to extract clinically meaningful insights from real-world practice without the prohibitive cost of large-scale trials. This does not replace trials but augments them, filling knowledge gaps that would otherwise persist.

The study ultimately points toward the future of predictive ophthalmology: a discipline where patients can receive individualized survival estimates incorporating their diagnosis, demographics, and surgical details, generated by models validated on thousands of prior cases. As eye banks, surgeons, and patients weigh their options, machine learning-driven predictions can become integral to shared decision-making. In this sense, the story of DSAEK versus PK is not only about surgical technique but about the maturation of a field that embraces computation as a partner in clinical judgment.

Study DOI: https://doi.org/10.3389/fmed.2022.831352

Engr. Dex Marco Tiu Guibelondo, B.Sc. Pharm, R.Ph., B.Sc. CpE

Editor-in-Chief, PharmaFEATURES