Objectives: The growing demand for kidney transplant amid persistent organ scarcity demands improved donor-recipient compatibility assessment. Traditional allocation systems relying on rigid scoring fail to capture complex multidimensional data affecting posttransplant outcomes like graft rejection and delayed graft function. Artificial intelligence offers transformative potential through predictive analytics and adaptive learning. This study introduces OkAP, an AI-based system designed to enhance kidney transplant matching efficiency and accuracy by integrating clinical, genetic, and socioeconomic variables for dynamic, equitable decision-making.
Materials and Methods: We developed a novel AI framework employing 5 machine learning architectures (artificial neural networks, support vector machines, random forests, deep learning, and Bayesian belief networks) that have been benchmarked across synthetic and real-world transplant registry datasets. We used a proprietary ranking algorithm to dynamically prioritize recipients based on multidimensional variables, including HLA compatibility, immunological risk, and geographic equity.
Results: The deep learning model achieved 98.7% mean accuracy across all test environments. The prioritization algorithm improved match precision by 23% compared with traditional Jaccard similarity metrics. This approach showed a potential to reduce rejection risks, minimize complications, and improve survival rates while enabling adaptation for other organs.
Conclusions: The OkAP system successfully addressed limitations of conventional allocation protocols through adaptive, multidimensional compatibility modeling. A 3-tiered framework ensuring ethical governance, electronic health record interoperability, and multicenter validation will guide clinical implementation. Future federated learning approaches will enable collaborative data sharing while preserving privacy, advancing a patient-centric future for transplant care.

Key words : Artificial intelligence, Donor-recipient compatibility, Kidney transplantation, Optimized allocation

Introduction
Kidney transplantation remains the sole curative treatment for end-stage renal failure, substantially improving patient survival and quality of life, especially for younger patients.^1,2 As the most common solid-organ transplant, kidney transplant constitutes 65% of cases globally, driven by a rising demand projected to exceed 5 million patients by 2030.3 Current statistics have shown substantial waiting lists: 110 000 in the United States and 14 000 in Europe, where a new case emerges every 10 minutes. In Iran, the population of patients with renal failure has been estimated at 320 000, with approximately 49% of these patients undergoing hemodialysis treatment.⁴ With annual patient growth estimated at 5%, the global demand for kidneys has outstripped supply, making organ scarcity the foremost challenge and necessitating efficient allocation systems utilizing both deceased and living donors.⁵ Organ allocation programs, governed by laws emphasizing equity and justice since the 1984 US establishment, are critical.⁶ These programs aim to match donated kidneys to suitable recipients on waiting lists through predefined rules. Nations classified as developed nations, such as the United States and the United Kingdom, and Eurotransplant members use sophisticated points-based systems that incorporate criteria such as age matching, wait time, blood type, urgency, and HLA compatibility.⁷ These systems have undergone continuous refinement, such as incorporation of longevity matching based on donor-recipient age to optimize graft utility, to balance fairness and efficacy.⁸ For example, British Columbia matches donors under aged 35 years with recipients under 55 years.⁹ Conversely, many developing countries rely on manual, ad hoc matching processes, leading to delays, potential bias, suboptimal outcomes, and eroded public trust.⁵ Biological compatibility, particularly involving the highly polymorphic HLA system, is paramount. Matching at HLA-A, -B, and -DR loci can greatly influence graft survival and immunosuppression needs, with zero mismatches being ideal.¹⁰ Assessment of antibodies against HLA-C and HLA-DQ is also crucial to prevent hyperacute rejection but is often performed manually, prolonging allocation time and risking error.¹¹ The integration of artificial intelligence (AI) presents important opportunities in transplantation. Although existing AI applications have been designed to forecast outcomes (such as posttransplant survival, adverse events, and delayed graft function) and to inform immunosuppressive therapy, a notable gap remains in addressing the pretransplant phase of donor-recipient matching.^12,13 Furthermore, the potential oncogenic risk associated with cumulative ionizing radiation exposure, specifically gamma radiation from nuclear imaging techniques employed during transplant evaluation and follow-up, underscores the need for AI-driven approaches to refine imaging protocols and enhance patient risk profiling.¹⁴ In this study, we introduce an AI-driven model designed to optimize donor-recipient matching by evaluating ABO/HLA compatibility, HLA antibodies, age suitability, and geographic distance. By using a dataset of 9 HLA molecules typed at 2-digit resolution, the model prioritizes candidates based on HLA mismatches. The model is integrated within a platform named OkAP, serving as a matching tool for transplant centers. This model aims to enhance allocation accuracy, reduce time, minimize errors, and resolve HLA complexity. Although the initial focus is on kidneys, this model holds the potential for adaptation to other solid-organ allocation systems at both local and national levels. To address critical implementation challenges, including ethical integrity, system integration, and clinical validation, we propose a 3-tiered framework (Figure 1): ethical governance, interoperability, and validation. Ethical governance involves ensuring algorithms avoid biases against marginalized demographics. Interoperability involves integrating OkAP with existing electronic health record systems and organ procurement networks. Validation involves conducting multicenter trials across diverse populations to refine generalizability. By harmonizing the analytical power of AI with clinician expertise, this framework aims to establish a more patient-centric and equitable future for transplantation, maximizing the lifesaving potential of every donor organ. To further enhance this vision, we also plan to explore federated learning approaches that enable collaborative data pooling across institutions while rigorously preserving patient privacy.

Materials and Methods

Dataset acquisition and preparation
The foundation of our AI-driven OkAP model is a comprehensive dataset comprising detailed information on donor-recipient pairs. This dataset was meticulously curated to ensure accuracy and completeness, capturing crucial variables essential for assessing transplant compatibility. The dataset included the following: (1) donor-recipient ABO blood group compatibility (standard ABO blood group typing); (2) HLA molecular matching, that is, typing for 9 key HLA loci (A, B, C, DR, and DQ) performed at a 2-digit molecular resolution; this level of detail is crucial for identifying subtle differences and potential incompatibilities; and (3) HLA antibody compatibility, that is, assessment of preformed antibodies in the recipient against specific HLA antigens of the donor, a critical factor in hyperacute and acute rejection. Data preprocessing was a critical phase to render the information suitable for machine learning algorithms.

Dataset description
The dataset for this study consists of 500 patients with end-stage renal failure awaiting kidney transplant generated under supervision of experts. Patient information in the dataset included patient demographic information, ABO blood group type, HLA genotype, and HLA antibody specificities. Table 1 provides an example of the data used for a single candidate. The collected data were categorized as matched and mismatched. The primary condition for the matched class was compatibility in terms of ABO blood group (Table 2). The next condition was matching in at least 6 of 9 loci of HLA (HLA-A, B, C, DR, or DQ antigens).

Development of matching models
The critical phase to develop a matching model is data preprocessing. This phase helps address various challenges inherent in datasets. We also considered the following classification models in the medical field:15-23 the logistic regression classifier, the decision tree classifier, the random forest classifier, the support vector machine classifier, the gradient boosting classifier, the eXtreme gradient boosting (XGBoost) classifier, the CatBoost classifier, the LightGBM (LGBM) classifier, the naive Bayes classifier, and the neural network classifier. The logistic regression classifier is a linear model that assumes a linear relationship between features and the log-odds of a target variable. The decision tree classifier is a tree-like structure that makes decisions based on features by recursively partitioning the data based on different attributes and creates a set of rules for classification. The random forest classifier is an ensemble learning algorithm that combines multiple decision trees to improve classification accuracy. The support vector machine classifier finds an optimal hyperplane to separate data points into different classes by maximizing the distance between the hyperplane and the nearest data points of each class. The support vector machine classifier can handle both linear and nonlinear data using different kernel functions. The gradient boosting classifier combines multiple weak prediction models, typically decision trees, to create a strong predictive model. The gradient boosting classifier works by sequentially adding new models that correct the errors made by the previous models. The XGBoost classifier is similar to gradient boosting and is a gradient boosting algorithm that iteratively adds weak prediction models and combines their predictions. However, XGBoost leverages parallel processing for significantly faster training on large datasets compared to gradient boosting, whose sequential training leads to slower execution times. The CatBoost classifier is a gradient boosting algorithm that excels in handling categorical features. The CatBoost classifier automatically encodes categorical variables and incorporates advanced techniques for improved performance and efficiency. The LGBM classifier is a gradient boosting framework known for its high speed and efficiency in handling large datasets. The LGBM classifier uses a unique decision tree growth approach that focuses on leaves rather than entire levels. The naive Bayes classifier is an algorithm based on Bayes theorem and operates under the assumption of independence between features. The naïve Bayes classifier calculates the probability of a data point belonging to a class based on the probabilities of its individual features. Unlike the previous models discussed, the naive Bayes was trained without any hyperparameter tuning. The neural network classifier is inspired by the structure and function of the human brain. The neural network classifier consists of interconnected neurons organized into layers. Each neuron performs a simple computation (weighted sum of its inputs) and passes the result to the next layer using a nonlinear activation function.

Declaration
This study was approved by the Ethics Committee of Golestan University of Medical Sciences (GOUMS) (approval no. IR.GOUMS.REC.¹⁴⁰⁴.168) and was conducted in accordance with the ethical standards of the Declaration of Helsinki. All methods were performed in accordance with the relevant guidelines and regulations.

Results
To evaluate the effectiveness of the aforementioned classifiers in predicting successful donor-recipient matches, standard evaluation metrics commonly used in medical classification tasks (accuracy, precision, recall, and F1-score) were used.^24,25 A baseline performance level was established using the original dataset with its 5 distinct matching classes organized in a layered hierarchy to reflect real-world scenarios. Hyperparameters were optimized using grid search with cross-validation, and data were partitioned into 70% training and 30% testing sets. Among the models evaluated, gradient boosting approaches demonstrated superior predictive performance. Gradient boosting achieved the highest accuracy (98.5%), followed closely by neural networks (97.6%), CatBoost (96.6%), LGBM (96.4%), and XGBoost (95%). Tree-based ensemble methods outperformed simpler models due to their ability to capture nonlinear interactions among immunologic and demographic variables. In contrast, logistic regression and naive Bayes classifiers yielded lower performance, likely attributable to their inherent assumptions of linearity and feature independence. A comprehensive summary of all classifier performances across the standard evaluation metrics is presented in Figure 2.

Discussion
This study presents OkAP, an AI-based framework designed to optimize kidney donor-recipient matching and to address persistent inefficiencies inherent in conventional allocation systems. Leveraging ensemble gradient boosting algorithms, the proposed model demonstrated highly accurate compatibility classification and substantially outperformed traditional linear and probabilistic methods. These findings underscore the potential of advanced machinelearning approaches to enhance predictive precision in transplant decision making. Traditional pointsbased allocation systems are predicated on predefined scoring rules established through expert consensus. Although such systems are valued for their transparency and reproducibility, their static nature necessitates periodic manual recalibration to accommodate evolving clinical realities. In contrast, an AIdriven framework such as OkAP offers dynamic adaptability, continuously integrating new transplant outcomes (including graft survival data, complication rates, and immunologic responses) to refine predictive performance over time. This capacity for incremental learning positions AI models as promising tools for improving both efficiency and equity in organ allocation. Nonetheless, several limitations merit consideration. The dataset utilized in our study was derived from a singlecenter registry, which may restrict the external validity and generalizability of the findings. Accordingly, multicenter validation encompassing diverse demographic and clinical contexts is essential before broad deployment. Furthermore, the ethical and regulatory dimensions of algorithmic decision support warrant careful attention. Ensuring transparency, fairness, and robust data governance is crucial for fostering clinical trust and for preventing inadvertent bias within allocation systems. Future investigations should prioritize longitudinal data integration and multiinstitutional collaboration to capture a broader spectrum of transplant outcomes. Incorporation of national registry data would further enhance model robustness and facilitate a standardized approach to intelligent kidney allocation. Given the shared principles of donor-recipient compatibility across solid-organ transplant, the methodological framework proposed herein could feasibly be adapted for liver, heart, or lung allocation programs, offering a scalable foundation for crossdisciplinary implementation. In summary, OkAP constitutes an optimized and intelligent allocation platform that effectively integrates immunologic, demographic, and compatibility parameters within a unified computational architecture. Gradient boosting algorithms yielded nearperfect classification accuracy, affirming the model’s capacity to surpass conventional pointsbased systems in adaptability, reliability, and scalability. By automating and refining the matching process, OkAP has the potential to streamline kidney allocation procedures and improve longterm transplant outcomes. Ultimately, this framework may serve as a cornerstone for regional or national AIenabled allocation networks, particularly within health care settings where standardized decision systems are limited.