Objectives: Organ transplantation functions as a critical medical procedure that serves patients with severe multiorgan conditions and chronic diseases. Early and accurate predictions of requirements for transplant enhance patient care through better clinical choices and improved resource allocation and treatment selection. We developed a complex artificial intelligence forecasting system to determine organ transplant needs, specifically kidney diseases by analyzing clinical data. Accurate and timely prioritization of patients for kidney transplant is a critical challenge due to organ scarcity and the complexity of clinical decision-making.
Materials and Methods: In this study, we propose an artificial intelligence-driven predictive framework to assess kidney transplant necessity by integrating patient demographics, clinical severity indicators, donor matching status, and real-time organ condition data. We analyzed a structured dataset consisting of 1000 kidney transplant records with 25 attributes by using ensemble learning models, including random forest and gradient boosting, and a deep neural network model. The prediction task was formulated as a binary classification problem reflecting real-world transplant outcomes.
Results: Experimental results demonstrated that the deep neural network consistently outperformed ensemble-based models, achieving superior accuracy, precision, recall, and F1-score with an area under the receiver operator characteristic curve of 0.95. Threshold-based sensitivity analysis confirmed that the deep learning model provided a clinically favorable trade-off between high sensitivity and acceptable specificity.
Conclusions: The proposed approach shows promising results in helping medical professionals detect organ transplant candidates through complex risk evaluation assessments.

Key words : Artificial intelligence, Clinical decision support system, Organ transplant prioritization

Introduction
Kidney transplant is widely recognized as the most effective treatment for patients with end-stage renal disease, offering improved survival rates and quality of life compared with long-term dialysis.¹ However, persistent shortages of donor organs, coupled with the increasing number of patients awaiting transplant, pose a major challenge to health care systems worldwide. As a result, accurate and timely prioritization of transplant candidates has become a critical and complex clinical task that requires balancing considerations of medical urgency, donor availability, and organ viability.² Traditional kidney transplant decision-making relies heavily on clinician expertise and predefined clinical criteria, such as disease severity, patient age, and donor-recipient compatibility.³ Although these approaches are clinically established, they may not fully capture the complex, nonlinear relationships among multiple patient- and organ-related factors, particularly when real-time organ condition data are involved.⁴ Moreover, subjective judgment and inter-clinician variability can further complicate consistent and transparent prioritization decisions.⁵ In recent years, artificial intelligence (AI) has demonstrated substantial potential in health care applications, including disease diagnosis, outcome prediction, and clinical decision support. In the context of organ transplantation, AI-based models offer the ability to integrate heterogeneous data sources and uncover hidden patterns that may not be apparent through conventional statistical methods. Several studies have explored the use of machine learning (ML) to predict graft survival, posttransplant complications, and patient outcomes.^6,7 However, studies that focused on the prediction of transplant necessity itself are scarce, especially studies that integrate patient characteristics, donor matching information, and real-time organ condition indicators.^8,9 Previous work has demonstrated the value of dynamic and interpretable ML models that can continuously update posttransplant risk predictions, often achieving superior discrimination compared with baseline-only or static approaches. At the same time, systematic reviews and comparative studies emphasize that ML does not universally outperform classical survival models, highlighting the importance of rigorous benchmarking, calibration, and clinical interpretability. More recent research has expanded AI applications to pretransplant decision support, including organ quality assessment and delayed graft function prediction, although comprehensive prioritization and allocation frameworks remain relatively underexplored. Table 1 summarizes key AI- and ML-based studies most closely related to this work, emphasizing their methodological approaches, data characteristics, and relevance to kidney transplant outcome prediction.^10-17 In this study, we addressed this gap by proposing an AI-driven predictive framework for assessing kidney transplant necessity. By leveraging ensemble learning models and a deep neural network model, our proposed approach aimed to capture complex interactions among clinical severity indicators, donor matching status, and organ viability metrics. Unlike previous work that primarily emphasized posttransplant outcomes, this study focused on pretransplant prioritization, which is a critical yet underexplored stage in the transplant workflow.¹⁸ The main contributions of this study are 3-fold. First, a comprehensive dataset comprising patient demographics, clinical severity markers, donor matching information, and real-time organ condition data was utilized to model transplant necessity as a binary classification problem. Second, ensemble learning models were systematically compared with a deep neural network model to evaluate their effectiveness in capturing nonlinear decision patterns. Third, the proposed framework was evaluated by using clinically relevant performance metrics and threshold-based sensitivity analysis, emphasizing patient safety and real-world applicability.

Materials and Methods
We used a retrospective, data-driven predictive modeling framework to investigate the feasibility of ML and deep learning techniques for prediction of organ transplant necessity in kidney transplant. The primary objective was to develop and compare ensemble-based ML models and deep neural networks capable of supporting clinical decision-making by accurately identifying patients with higher transplant priority based on multidimensional patient donor data. Figure 1 shows the flowchart of the methodology process in this study. The problem was formulated as a binary classification task, in which the model predicted whether a patient proceeds to kidney transplant based on clinical, biological, and real-time organ indicators.

Dataset description
We obtained the dataset used in this study from the publicly accessible dataset, entitled “Kidney Organ Transplantation Patient-Donor Data.”19 The dataset comprises 1000 anonymized patient-level records described by 25 attributes that collectively capture key factors influencing kidney transplant decision-making. The dataset encompasses a diverse set of variables, including patient demographic information such as age, body weight, body mass index, and blood type, along with clinical diagnosis outcomes corresponding to different stages of chronic kidney disease. Disease severity is further represented through quantitative biological markers reflecting renal dysfunction. In addition to patient-related information, the dataset contains donor eligibility constraints and patient-donor matching status to reflect practical transplant conditions. Notably, real-time organ condition indicators are also included, such as a continuous organ health score and categorical alert levels, which characterize dynamic organ viability during the transplant process. An AI-generated survival probability estimate was also provided, serving as a complementary prognostic indicator alongside clinical variables. All records were fully anonymized and free of personally identifiable information, ensuring suitability for research use. For supervised learning purposes, transplant outcome was formulated as a binary target variable derived from organ status, in which cases labeled as transplanted represented confirmed transplant necessity and those labeled as matched or pending constituted those not transplanted. This multidimensional data structure enabled comprehensive modeling of kidney transplant necessity, supporting development and evaluation of ML and deep learning-based clinical decision support systems.

Data preprocessing
Before model development, several preprocessing steps were applied to ensure data consistency and algorithm compatibility. Categorical variables, including patient blood type, diagnosis category, matching status, and organ condition alerts, were converted into numerical representations using label encoding. Continuous variables, such as patient physical measurements, biological markers, real-time organ health scores, and predicted survival chances, were standardized using z-score normalization. To avoid information leakage, data scaling procedures were applied only after splitting the dataset into training and testing subsets. Feature selection was guided by clinical relevance and domain expertise rather than automated dimensionality reduction techniques. The selected feature set included patient demographic indicators, disease severity markers, real-time organ condition metrics, matching-related variables, and AI-derived survival probability estimates (Table 2). The dataset was partitioned into training and testing subsets by using an 80:20 ratio with stratified sampling applied to preserve the original class distribution. Model training was conducted exclusively on the training subset, whereas the testing subset was reserved for independent evaluation. This separation ensured an unbiased assessment of model generalization performance and reflected real-world deployment scenarios in which predictions are made on previously unseen data.

Ensemble learning models
Two ensemble-based ML models were implemented to capture nonlinear relationships and feature interactions within the dataset. The random forest classifier employed an ensemble of decision trees constructed through bootstrap aggregation, enabling robust performance and resistance to overfitting. In parallel, the gradient boosting classifier was used to sequentially optimize prediction accuracy by focusing on previously misclassified samples. These models were particularly suitable for structured medical data and provided a strong baseline for comparison with deep learning approaches. As shown in Figure 2, ensemble-based ML models could capture complex, nonlinear relationships within the kidney transplant dataset, with random forest and gradient boosting providing complementary predictive perspectives. Multidimensional input features derived from patient demographics, clinical severity indicators, donor matching status, and real-time organ condition metrics were processed in parallel by 2 ensemble learning models. The random forest classifier leveraged bootstrap aggregation of multiple decision trees to capture nonlinear feature interactions and improve robustness against overfitting, and the gradient boosting classifier sequentially refined predictions by emphasizing previously misclassified cases. The outputs of both models contributed to the final prediction of kidney transplant necessity.

Deep neural network architecture
In addition to ensemble models, a deep learning approach based on a multilayer perceptron (MLP) architecture was developed to model complex and high-dimensional feature interactions. The deep neural network consisted of multiple fully connected hidden layers with rectified linear unit activation functions to introduce nonlinearity. Regularization mechanisms were incorporated to mitigate overfitting, and a sigmoid activation function was used in the output layer to generate probabilistic predictions for binary classification. This architecture enabled the model to learn hierarchical feature representations relevant to transplantation necessity (Figure 3).

Performance evaluation metrics
We assessed model performance by using multiple evaluation metrics, including accuracy, precision, recall, and F1-score. We placed special emphasis on recall, given its clinical importance in minimizing false-negative predictions that could result in delayed or missed transplant for high-risk patients. Precision was also considered to avoid unnecessary prioritization of low-risk cases, thereby supporting balanced and clinically responsible decision-making. We conducted all experiments with established Python-based ML libraries to ensure reproducibility and methodological transparency. Ethical considerations were integral to the study design, and the dataset was fully anonymized and contained no personally identifiable information. The proposed AI framework was intended solely as a clinical decision support tool and was not intended to replace physician judgment, aligning with ethical guidelines for responsible use of AI in health care.

Results and Discussion
The experimental findings were obtained from the ensemble learning and deep learning models that were developed for transplant necessity prediction. The evaluation focused on overall predictive performance, class-wise discrimination capability, and the relative importance of key clinical and contextual features. The comparative performance results of the random forest, gradient boosting, and deep neural network MLP models are summarized in Table 3. Among the evaluated approaches, the MLP achieved the highest overall accuracy of 92.3%, outperforming both random forest (89.1%) and gradient boosting (87.6%). This superiority was consistently reflected across all evaluation metrics, including precision (90.8%), recall (94.1%), and F1-score (92.4%). The higher recall value of the MLP model was particularly significant in the context of transplant necessity prediction, where failing to identify patients who genuinely require transplant may have critical clinical consequences. The random forest model demonstrated a balanced and robust performance, with strong recall (91.2%) and an F1-score of 89.2%, indicating its effectiveness in capturing complex, nonlinear relationships within the data. Gradient boosting, although competitive, exhibited comparatively lower performance across all metrics, suggesting a higher sensitivity to data variability or feature interactions in this specific application. Table 4 provides a detailed class-wise analysis of prediction performance for the 2 target classes: transplant needed and transplant not needed. Across all models, performance metrics were consistently higher for the transplant needed class, which is desirable given the clinical priority of correctly identifying high-risk patients. The deep neural network achieved the best class-wise performance, with a precision of 91.5%, recall of 94.1%, and F1-score of 92.8% for the transplant needed class. This indicated that the model not only accurately identified patients requiring transplant but also minimized false positives. For the not needed class, the MLP maintained strong discrimination capability, achieving an F1-score of 90.3%, which demonstrated balanced decision-making across both classes. The random forest model also showed reliable class-wise behavior, particularly with a recall of 91.2% for the transplant needed class, suggesting its suitability for risk-sensitive decision support systems. The gradient boosting model, although effective, exhibited lower precision and recall values for both classes, indicating comparatively weaker generalization in distinguishing borderline cases. Our results highlighted the advantage of deep learning models in handling class boundaries and overlapping feature distributions, which are common challenges in medical decision-making datasets. Table 5 lists the most influential predictive features identified by the ensemble models. Both the random forest and gradient boosting models consistently ranked “RealTime_Organ_HealthScore” and “Predicted_Survival_Chance” as the most critical features, underscoring the importance of dynamic physiological indicators and prognostic estimates in transplant decision processes. The features “Biological_Markers” and “Match_Status” also showed substantial importance across both models, reflecting the clinical relevance of biochemical indicators and donor-recipient compatibility. Features such as “Diagnosis_Result,” “Patient_Age,” and “Patient_BMI” contributed moderately to the prediction outcomes, suggesting that, although demographic and diagnostic factors remain relevant, real-time and prognostic features play a more decisive role. The similarity in feature importance rankings between random forest and gradient boosting enhanced the interpretability and reliability of the results, indicating consistent learning behavior across ensemble techniques. Moreover, the strong performance of the deep neural network, despite not explicitly providing feature importance scores, suggested that this tool effectively integrated the key predictors in a nonlinear and hierarchical manner. Receiver operating characteristic comparing the random forest, gradient boosting and deep neural network models The deep neural network achieved the highest discriminative performance with an area under the curve (AUC) of 0.95, followed by the random forest (AUC = 0.91) and the gradient boosting (AUC = 0.89) models. Random classification is shown in Figure 4. In kidney transplant, false-negative predictions that classified a high-priority patient as low priority may result in delayed or missed transplant and potentially severe clinical consequences. Therefore, beyond global performance metrics, a threshold-based sensitivity analysis was conducted to evaluate model behavior under clinically conservative decision thresholds. This analysis focused on optimizing sensitivity (recall) while maintaining acceptable specificity, reflecting real-world transplant prioritization strategies. The probabilistic outputs generated by each model were evaluated across multiple decision thresholds, with particular emphasis on thresholds that achieved high sensitivity levels (≥90% and ≥95%), as recommended for safety-critical clinical decision support systems (Table 6). At sensitivity levels exceeding 90%, all models demonstrated the ability to identify most high-priority transplant candidates. However, notable differences emerged in the trade-off between sensitivity and specificity. The deep neural network consistently maintained higher specificity and precision at equivalent sensitivity levels compared with ensemble learning models. At a sensitivity target of 95%, the deep neural network achieved a specificity of 79.8%, significantly outperforming the random forest and gradient boosting models, which experienced substantial specificity degradation. The threshold-based sensitivity analysis demonstrated that the deep neural network offered superior robustness under conservative operating conditions. Even when configured to aggressively minimize false negatives, the deep neural network preserved a substantially higher specificity compared with ensemble learning models. This property is particularly critical in kidney transplant workflows, where prioritizing patient safety must not excessively compromise organ allocation efficiency (Table 7). Our findings indicated that the proposed deep learning model was better suited for deployment as a clinical decision support system operating under high-sensitivity constraints, ensuring that patients with urgent transplant needs are rarely overlooked while maintaining controlled false-positive rates.

Conclusions
Deep learning-based models can significantly enhance the accuracy and reliability of kidney transplant prioritization. By serving as a clinical decision support tool, the proposed framework has the potential to assist clinicians in making more informed, consistent and transparent transplantation decisions, ultimately contributing to safer and more efficient organ allocation processes. The AI-driven predictive framework in our study assessed kidney transplant necessity by integrating patient demographics, clinical severity indicators, donor matching status, and real-time organ condition data. Ensemble learning models and a deep neural network were systematically evaluated to capture complex, nonlinear relationships governing transplant prioritization. Our study showed that, although ensemble-based approaches provided strong and interpretable baseline performance, the deep neural network consistently achieved superior predictive accuracy, recall, F1-score, and discriminative capability. Threshold-based sensitivity analysis further confirmed that the deep learning model maintained a favorable balance between high sensitivity and acceptable specificity, which is critical for minimizing false-negative decisions in kidney transplant workflows. Despite promising outcomes, our study had several limitations. The dataset did not include immunological compatibility variables such as HLA typing, crossmatch results, or genomic information, which are important determinants of transplant success. In addition, the analysis was limited to transplant occurrence and did not incorporate longitudinal posttransplant outcomes such as graft survival or patient mortality. The use of a single structured dataset may also limit the generalizability of the findings across different clinical centers and populations. Furthermore, although the deep neural network exhibited superior performance, its reduced interpretability compared with ensemble models may pose challenges for clinical adoption. Future research is recommended to incorporate multicenter and multimodal datasets, integrate explainable AI techniques to enhance transparency, and perform prospective validation within real-world clinical decision support systems.