AI and Machine Learning in Transplantation

Abstract

Artificial Intelligence (AI) and machine learning (ML) are increasingly being applied across the transplantation care pathway, supporting tasks such as donor–recipient matching, immunological risk stratification, early detection of graft dysfunction, and optimisation of immunosuppressive therapy. This review provides a structured synthesis of current AI applications in transplantation, with a focus on underrepresented areas including real-time graft viability assessment, adaptive immunosuppression, and cross-organ immune modelling. The review also examines the translational infrastructure needed for clinical implementation, such as federated learning, explainable AI (XAI), and data governance. Evidence suggests that AI-based models can improve predictive accuracy and clinical decision support when compared to conventional approaches. However, limitations related to data quality, algorithmic bias, model transparency, and integration into clinical workflows remain. Addressing these challenges through rigorous validation, ethical oversight, and interdisciplinary collaboration will be necessary to support the safe and effective use of AI in transplant medicine.

1. Introduction

1.1. Clinical Challenges in Transplantation

Transplantation has evolved significantly over recent decades, driven by advancements in surgical techniques, immunosuppressive regimens, and postoperative management. However, major challenges persist. Equitable organ allocation remains a pressing issue, with demand consistently outpacing supply. Graft survival, though improved, is still vulnerable to immune rejection, delayed recovery, and post-transplant complications. Clinicians must navigate complex trade-offs, particularly in the early post-transplant period, such as adjusting immunosuppression to prevent rejection while minimising the risk of opportunistic infections and malignancies. These multifaceted clinical demands underscore the urgent need for novel strategies to enhance decision-making, optimise resource use, and personalise care pathways across the transplant continuum.

1.2. Technological Potential of AI and ML

Artificial Intelligence (AI) and machine learning (ML) offer transformative opportunities for addressing the intricacies of transplant medicine. These technologies can process large-scale, multidimensional data—including clinical, imaging, molecular, and physiological inputs—to support more informed, real-time decision-making. ML models have shown potential in predicting post-transplant complications, stratifying immunological risk, tailoring immunosuppressive therapy, and identifying early signs of graft dysfunction. AI-driven tools are increasingly being deployed to refine organ allocation algorithms and improve donor–recipient matching by leveraging pattern recognition that surpasses human capability. Decision support systems powered by AI may significantly reduce variability in clinical practice, ultimately improving patient outcomes and care efficiency.

1.3. Review Scope and Novel Contribution

This review presents a longitudinal synthesis of AI applications across the full transplantation care pathway—from preoperative risk assessment and intraoperative monitoring to postoperative rehabilitation and long-term outcome prediction. In contrast to existing literature, we place special emphasis on underrepresented domains such as real-time graft viability assessment, adaptive immunosuppression management, and immune modelling across different organ systems. Moreover, we address the translational infrastructure required to realise AI’s clinical potential, including federated learning architectures, explainable AI, and robust data governance models. By integrating technical, clinical, and organisational perspectives, this review aims to map the evolving role of AI in transplant medicine and identify actionable opportunities for future development.

2. AI and ML in Immunological Assessment

2.1. Pre-Transplant Compatibility

Advancements in AI and ML have significantly improved the prediction of histo-compatibility and crossmatch outcomes in transplantation. Traditional methods for HLA matching and crossmatch prediction rely on serological and molecular techniques that, while effective, have limitations in sensitivity and specificity. AI-driven approaches integrate multi-omics data, transcriptomic profiling, and machine learning algorithms to enhance prediction accuracy [1]. These tools include models like the UK Live-Donor Kidney Transplant Outcome Prediction tool, which utilises AI to improve prognostication and guide donor–recipient selection [2,3,4,5]. The PROMAD atlas, for instance, employs a pan-organ transcriptomic model that standardises gene expression data across multiple organ transplants, identifying shared immune signatures associated with rejection risk [1].

ML models also facilitate the identification of senescence-related genes (SRGs) that influence transplant rejection. Shen et al. demonstrated that pre-transplant senescent cells, particularly those expressing p21Cip1 and p16INK4a, correlate with interstitial fibrosis and chronic allograft nephropathy [6]. Their ML-based approach enabled the identification of 13 predictive SRGs, highlighting their potential for non-invasive compatibility assessment [6]. Furthermore, Bayesian AI models, such as Bayes-CRE, enhance predictive robustness by integrating causal biological networks with transcriptional profiles, reducing the risk of overfitting [7]. These AI-driven methodologies offer superior precision in risk assessment, thereby improving donor–recipient matching and long-term graft survival (Appendix A).

2.2. Risk Stratification

The ability to stratify transplant recipients based on their risk of immune-mediated rejection is a critical component of post-transplant management. ML models utilise biomarkers, transcriptomic data, and historical clinical cases to predict rejection risk more accurately than conventional methods. Emerging ML platforms, such as those described by Weimer and Newhall, incorporate multi-factor immunological features to provide dynamic risk stratification with clinical applicability [8]. The integration of high-throughput omics technologies, including single-cell RNA sequencing and proteomics, has facilitated the identification of novel rejection biomarkers [9]. TruGraf and kSORT, commercially available blood-based gene expression tests, exemplify AI-assisted tools that identify immunological quiescence, enabling early rejection detection and minimising the need for protocol biopsies [9].

Shen et al. applied ML techniques to renal biopsy transcriptomes, identifying TNF and NF-κB pathway involvement in rejection [6]. Shao et al. further substantiated the role of these pathways in T-cell-mediated liver allograft rejection through integrative RNA-seq and ML approaches [10]. Their model incorporated IL2RB, IL15, TNFSF13B, and C1QA as predictive markers, achieving an AUC of 97.5% for rejection detection [6]. Similarly, Robertson et al. identified myeloid cell-derived transcriptomic markers common to multiple transplanted organs, suggesting that pan-organ immune risk stratification could refine post-transplant surveillance strategies [1]. Moreover, ML-driven models are being developed to predict long-term graft dysfunction based on early biopsy samples, with applications extending to normothermic machine perfusion (NMP) interventions [1].

Liver transplantation presents unique challenges in immune risk stratification due to its tolerogenic properties. AI-based models are increasingly being applied to monitor operational tolerance and guide immunosuppression weaning (ISW). Studies have identified CXCL10 and FOXP3 as early markers of rejection in ISW failure cases, with miRNA-based models achieving 88.9% sensitivity and 83.3% specificity in predicting rejection prior to histological evidence [11]. These findings underscore the role of AI in refining immunosuppressive strategies and minimising rejection risk (Appendix A).

2.3. Tolerance Modelling

Autoimmune response modelling plays a pivotal role in assessing recipient immune tolerance and predicting long-term transplant outcomes. AI-driven tools are being utilised to decipher immune regulatory networks and identify key determinants of transplant tolerance. Liver transplantation, due to its intrinsic tolerogenicity, provides a model for studying alloimmune response modulation [11]. AI-integrated transcriptomic analyses have revealed unique B cell-associated gene signatures in operationally tolerant kidney transplant recipients, including TCL1A and a composite 20-gene profile indicative of immunological stability [9]. Fu et al. similarly identified predictive gene sets associated with operational tolerance using an unbiased ML exploration, reinforcing the utility of AI in tolerance assessment [12].

Computational models leveraging AI are also being developed to integrate donor-derived cell-free DNA (dd-cfDNA) and extracellular vesicle (EV) data to non-invasively monitor immune tolerance [9]. These models aim to reduce reliance on invasive biopsies and improve real-time immune surveillance. Furthermore, unsupervised learning algorithms have classified transplant glomerulopathy into five distinct subtypes, each with varying graft survival rates (88–22%), highlighting the utility of AI in refining disease classification [9]. Peloso et al. have also reviewed the future potential of such AI-driven classification systems across various organ transplant domains [13].

The integration of causal inference frameworks with Bayesian AI models has further enhanced the predictive power of immune tolerance modelling [7]. These models distinguish between diagnostic signatures of disease progression and drug response, improving precision medicine applications in transplantation. Future advancements in AI-driven immune response modelling are expected to optimise patient stratification, facilitate personalised immunosuppressive regimens, and ultimately improve long-term transplant outcomes.

A summary of the key findings related to AI and ML in immunological assessment and areas for future research can be found in

Table 1. AI Applications in Transplant Immunology: Summary of AI tools for donor–recipient matching, early rejection detection, and immune profiling, alongside recommendations for Bayesian model scalability, multi-omics harmonisation, and integration of AI-driven decision support in immunosuppression management.

Key Findings

The UK Live-Donor Prediction tool, PROMAD atlas, and Bayes-CRE model improve donor–recipient compatibility and predict senescence-related gene activity linked to graft outcomes. TruGraf, kSORT, and models incorporating TNFSF13B, IL2RB, and myeloid markers allow early rejection detection and enable dynamic monitoring post-transplant across organ types. Unsupervised learning and Bayesian inference classify disease subtypes and distinguish immune response dynamics, supporting non-invasive monitoring and personalised immunosuppression.

Recommendations

Investigate scalable Bayesian frameworks for model generalisability, cross-organ applicability, and real-world integration into clinical workflows. Develop methodologies for the standardisation and harmonisation of multi-omics datasets. Evaluate the impact of AI-driven decision-support tools on immunosuppression weaning protocols.

.

3. AI in Designing Clinical Pathways

3.1. Decision-Support Systems

AI-based decision-support systems (DSS) are transforming patient selection and pre-transplant assessments by integrating vast datasets, including clinical, genetic, and imaging data. Traditional approaches to candidate selection often rely on clinician judgement and established scoring systems such as the Model for End-Stage Liver Disease (MELD) and Kidney Donor Risk Index (KDRI). However, AI-driven models offer enhanced risk stratification by identifying subtle patterns beyond human perception [14]. AI-enhanced DSS, such as those evaluated in qualitative studies on fairness and usability, are being increasingly explored for liver transplant candidate evaluation [15].

ML algorithms have demonstrated superior predictive capabilities for perioperative risk assessment in transplantation. Molinari et al. developed an ML-based predictive model for perioperative mortality in liver transplant candidates, achieving high discriminatory power (C-statistics of 0.95 for ≥10% mortality risk) [14]. These models assist in decision-making by providing objective risk estimates and supplementing traditional clinical evaluations. Emerging bibliometric analyses have also highlighted a surge in such AI applications over the past decade [16].

Despite their promise, challenges persist. Explainable AI (XAI) remains a major concern, as black-box models may introduce biases, limit interpretability, and complicate clinical adoption [17]. Consequently, efforts are directed towards integrating AI models that balance predictive performance with interpretability, such as logistic regression and rule-based models that retain clinical transparency while incorporating AI-driven insights. This balance has been extensively reviewed in the context of kidney transplantation, underscoring the need for interpretable yet accurate systems [17].

3.2. Predictive Modelling for Outcomes

Predictive modelling is pivotal in forecasting post-transplantation outcomes, including graft survival and complications. ML approaches leveraging gene expression profiles have enhanced our understanding of allograft tolerance and rejection mechanisms. Fu et al. employed an ML framework combining datasets from multiple studies (GSE166865, IOT, ITN) to develop a robust model for predicting kidney allograft tolerance, achieving 91.7% sensitivity and 93.8% specificity [12].

ML models have been particularly effective in identifying gene expression signatures associated with immune tolerance. Fu et al. highlighted five key genes—HLA-DOA, TCL1A, EBF1, CD79B, and PNOC—as crucial markers for allograft tolerance, with B-cell-related genes playing a central role [12]. The application of these biomarkers in AI-driven models could enable early identification of patients with a lower risk of rejection, facilitating personalised immunosuppression withdrawal strategies.

However, the application of ML in outcome prediction is not without limitations. Explainability issues persist, with concerns about adversarial attacks and biases in feature selection [17]. The choice of ML models also significantly impacts reliability; traditional regression models often outperform complex ML algorithms in structured datasets, such as the iBox model for kidney transplantation, which remains the gold standard in clinical practice [17]. Studies such as those by Salaün et al. further demonstrate how interpretable ML models can offer clinically relevant insights while retaining predictive robustness [18].

3.3. Optimisation of Follow-Up Schedules

Post-transplant care is highly individualised, requiring dynamic monitoring and adaptation. AI-driven approaches are increasingly being applied to optimise follow-up schedules by integrating real-time patient data from electronic health records (EHRs), biomarker analyses, and wearable technologies.

Predictive analytics can tailor follow-up intervals based on patient-specific risk factors, reducing unnecessary clinical visits while ensuring timely interventions for high-risk patients. For instance, ML models incorporating perioperative variables have demonstrated the ability to predict 90-day, 1-year, and 5-year survival in liver transplant recipients with high accuracy [14]. Recent liver transplant-focused innovations highlight how big data tools combined with AI can enhance post-operative monitoring [19].

Furthermore, AI can facilitate proactive interventions by detecting early signs of complications such as graft dysfunction or opportunistic infections. The integration of molecular biomarkers, as highlighted by Fu et al., could refine AI models to predict immunological tolerance, guiding decisions on immunosuppression tapering and minimising adverse effects [12]. However, challenges such as data privacy concerns, algorithmic biases, and the need for real-world validation remain significant barriers to clinical implementation [17]. Ethical frameworks and regulatory considerations for AI in transplant pathways are also being increasingly recognised [20].

A summary of the key findings related to AI in designing clinical pathways and areas for future research can be found in

Table 2. AI-Enhanced Post-Transplant Monitoring and Prognostics: Overview of AI models for mortality prediction, follow-up optimisation, and personalised care, highlighting the need for explainability, ethical data integration, and standardised validation across clinical settings.

Key Findings

ML models provide high-accuracy mortality risk estimates. While complex ML models offer strong performance, interpretable models like iBox and those emphasising explainability are more clinically reliable. Electronic health records, wearable data, and biomarkers enable early intervention and tailored immunosuppression strategies, though challenges like data privacy, algorithmic bias, and ethical regulation continue to hinder widespread implementation.

Recommendations

Develop methods to enhance explainability in high-performing black-box AI models used in transplantation. Design and validate robust frameworks for evaluating AI-driven follow-up protocols across diverse clinical environments. Establish ethical and secure approaches for integrating patient-generated data from wearable and digital devices into post-transplant AI systems.

.

4. AI in Detection and Treatment of Graft Dysfunction

4.1. AI-Based Diagnostic Tools for Early Graft Rejection

Accurate and early detection of graft rejection is critical for successful transplant outcomes. AI models have demonstrated their capacity to outperform traditional diagnostic approaches by integrating histopathological, radiological, and molecular data. Furness et al. showed that Bayesian belief networks could outperform expert pathologists in detecting acute renal allograft rejection, leveraging data fusion from laboratory and biopsy inputs [21]. Despite their promise, these models face implementation barriers such as inter-institutional variability and labour-intensive data curation.

In digital pathology, AI-enhanced platforms have shown utility in accelerating and standardising workflows. Baxi et al. demonstrated that AI-driven tools could enhance translational diagnostics in real time [22], while Seraphin et al. applied self-supervised deep learning to histological slides to predict heart transplant rejection, reducing observer variability across organ systems [23]. At the molecular level, Anglicheau et al. and Zhang et al. identified microRNA and lncRNA signatures, respectively, that can distinguish between rejection and stable graft function, underscoring their value as inputs for AI-based diagnostic algorithms [24,25]. Sharaby further proposed an AI pipeline to integrate multi-omics biomarkers into rejection diagnosis, potentially enabling more holistic, automated assessments [26].

Studies exploring inflammatory protein networks—such as ICAM-1, RANTES, and TIMP-2—also indicate promising roles in identifying chronic rejection, as suggested by Zhu et al. [27]. Moreover, mechanistic insights from xenotransplantation research by Park et al. reveal calcium-handling proteins as early rejection markers, adding to the spectrum of candidates for AI integration [28]. These findings are reinforced by reviews from Belčič Mikič and Arnol and Mizera et al., which collectively highlight the growing reliability of ML-based rejection prediction models across transplantation contexts [29,30].

4.2. Remote and Continuous Monitoring Through Digital Health Technologies

The convergence of AI and wearable technologies offers transformative potential for remote, real-time monitoring of graft health. Zhi et al. introduced RtNet, a model that integrates deep learning with multiparametric MRI to noninvasively assess renal allograft status. Their results revealed superior accuracy in distinguishing between acute rejection, chronic allograft nephropathy, and stable graft function compared to conventional MRI [31]. Buscher et al. contributed to this evolving paradigm by applying unsupervised learning for transplant phenotyping, identifying distinct immune states with differing prognostic implications [32]. This approach could refine personalised monitoring strategies through future applications of spatial transcriptomics and single-cell analyses.

In terms of point-of-care diagnostics, Lee et al. combined surface-enhanced Raman spectroscopy (SERS) with ML to achieve diagnostic accuracies of 93.53–98.82% in detecting biomarkers associated with renal allograft rejection, indicating feasibility for future portable or wearable devices [33]. Madhvapathy et al. advanced this further with implantable bioelectronic sensors enabling continuous biosensing for kidney rejection, while Keating et al. validated commercially available wearables for detecting physiological changes linked to infection and graft dysfunction [34,35].

Systematic reviews and expert commentaries—by Kuang et al., Falevai and Hassandoust, and Murray et al.—emphasise the growing evidence base for digital monitoring. The Reboot 2.0 trial underscores the clinical viability of mobile monitoring protocols to detect early deterioration in transplant recipients [36,37,38]. These developments demonstrate how AI-enabled digital health infrastructure can support timely intervention and reduce dependence on episodic, clinic-based assessments.

4.3. Personalising Immunosuppression Through AI-Driven Therapies

Tailoring immunosuppressive therapy to individual immune profiles is increasingly possible with AI models that incorporate clinical, genomic, and proteomic data. In liver transplantation, Pérez-Sanz et al. identified non-invasive biomarkers such as SENP6 and FEM1C that may predict immune tolerance, which could inform AI tools for adaptive dosing strategies [39]. Pei et al. identified apolipoprotein D and GPX3 as biomarkers of oxidative stress and immune activation in kidney transplantation, offering additional candidates for inclusion in predictive models that balance rejection prevention with toxicity risk [40,41].

Wang and Peng employed bioinformatic pipelines to identify biomarkers like ALAS2, HBD, EPB42, and FECH in heart transplant recipients, also noting the relevance of Th17-mediated responses. These insights could feed into AI models for dynamic immunosuppression adjustment [42]. A more integrative approach was proposed by Basuli and Roy, who developed a dynamic framework that adjusts therapy based on continuous patient-level biomarker data, potentially improving long-term graft outcomes while minimising over-immunosuppression [43].

Broader perspectives from Al Moussawy et al., Gotlieb et al., and Olawade et al. consolidate the landscape of AI-enabled immunosuppression optimisation across transplant types, advocating for smarter regimen selection and tolerance prediction [44,45,46]. Ramalhete et al. push this further by advocating for explainable AI and federated learning approaches to enhance personalisation while preserving data privacy—crucial for cross-institutional collaborations [47]. Collectively, these models point toward a future in which immunosuppressive care is continuously responsive to each patient’s evolving immunological status.

A summary of the key findings related to AI in detection and treatment of graft dysfunction and areas for future research can be found in

Table 3. AI Tools for Rejection Detection and Immunosuppression Management: Highlights AI-driven approaches for non-invasive rejection monitoring, early detection via imaging and biosensors, and personalised immunosuppression, with recommendations for model validation, wearable integration, and interpretability.

Key Findings

Integration of microRNAs, lncRNAs, and inflammatory protein markers into AI models allows non-invasive, multi-omics-based rejection prediction, reducing the reliance on biopsies. RtNet MRI, SERS diagnostics, and implantable biosensors detect early physiological or molecular changes linked to rejection, enabling personalised follow-up and early intervention. ML optimises immunosuppressive therapy by integrating molecular biomarkers (e.g., SENP6, GPX3, ALAS2) and clinical parameters for personalised dosing.

Recommendations

Develop strategies to improve the cross-institutional generalisability of AI models for graft rejection detection. Design clinical validation protocols for integrating AI-driven data from wearable technologies into real-time transplant monitoring workflows. Investigate the trade-offs between AI model complexity and interpretability in immunosuppressive therapy management.

.

5. Precise and Tailored Interventions

5.1. AI-Driven Optimisation of Immunosuppressive Therapy

Pharmacogenomics has emerged as a key domain for precision medicine in transplantation, where AI is enabling the personalisation of immunosuppressive drug regimens. Tang et al. demonstrated the utility of machine learning models to predict stable tacrolimus dosing in renal transplant recipients, identifying the CYP3A5 genotype—particularly the SNP 6986 A > G variant—as a crucial determinant of dosing requirements [48]. While the model performed well, especially in mid-dose ranges, it also highlighted the complexity of integrating genetic-environmental interactions, such as comorbidities and lifestyle factors, into pharmacological predictions.

Beyond genotyping, longitudinal data integration can enhance the predictive accuracy of dosing strategies. Moscatelli et al. showed how AI systems that consolidate data from laboratory tests, biopsies, and treatment histories could enable dynamic drug adjustments over time [49]. Similarly, Quaglia et al. demonstrated that combining multi-omics data, including biomarkers for immune activation and rejection, with AI models allows more granular tailoring of immunosuppression [9]. These advancements align with the broader precision medicine movement, as discussed by Nobakht et al. and Naesens and Anglicheau, who advocate for actionable biomarker-informed strategies that move beyond traditional risk stratification [50,51].

5.2. AI-Augmented Surgical Precision and Robotics

AI is reshaping surgical techniques in transplantation, particularly through robotic systems that enhance procedural accuracy. Gong et al. outlined the transformative role of AI-powered robotics in solid organ transplantation, highlighting their effectiveness in tasks such as tissue dissection and graft placement, which are critical for long-term graft survival [52]. These systems reduce human error and standardise complex procedures, although adoption is limited by high costs, technological variability, and clinical inertia.

Institutional readiness and regulatory frameworks remain crucial for successful integration. Johnston-Webber et al. emphasised the importance of governance structures in supporting AI technologies within transplantation ecosystems [53]. Ethical concerns—ranging from data privacy to informed consent—have been raised, as noted by Murdoch et al., who also stressed the value of clear communication and patient engagement when deploying AI tools in clinical settings [54]. The development of technologies such as explainable AI (XAI), federated learning, and blockchain further addresses transparency, trust, and interoperability issues, supporting future expansion of AI in surgery.

5.3. Personalised Rehabilitation Strategies Using AI

The use of AI in post-transplant rehabilitation is expanding beyond traditional monitoring to include predictive and adaptive recovery plans. Shung and Sung reported that AI systems can enhance clinical decision-making by identifying patients at risk for complications and adjusting rehabilitation protocols accordingly [55]. These models integrate patient-reported outcomes, biomarkers, and imaging data to offer a holistic assessment of recovery.

In practice, AI has facilitated structured rehabilitation programmes. Serper et al. showed that AI-guided behavioural interventions, such as personalised walking regimens, improved recovery outcomes in abdominal organ transplant recipients [56]. Quaglia et al. supported this by integrating biomarker-informed predictions into immunosuppression management, enabling early interventions and minimising unnecessary treatments [9].

AI also plays a role in stratifying patients for targeted rehabilitation. By incorporating proteomics and transcriptomics, AI models can anticipate post-transplant complications before clinical symptoms arise, allowing for precision-guided therapy adjustments. However, challenges such as data bias, inconsistency, and incomplete patient profiles must be addressed to ensure the accuracy and reliability of these tools. As Shung and Sung note, maintaining data quality and ongoing model retraining is essential for sustained effectiveness in rehabilitation contexts [55].

5.4. Policy and Governance Frameworks for AI-Driven Transplantation

For AI innovations to be successfully implemented in transplantation, they must be supported by robust policy and governance frameworks. These systems need to ensure data interoperability, organ availability, and equitable access to AI-powered tools. The KDIGO Clinical Practice Guidelines offer foundational recommendations for candidate selection, donor evaluation, and pre-transplant optimisation, which align with AI’s capabilities for risk assessment and outcome prediction [57].

National initiatives, such as the UK Organ Utilisation Group’s recommendations, underscore the role of AI in improving organ allocation and utilisation efficiency, linking technological advancement with public health infrastructure [58]. Philosophical contributions from Naesens propose a model of clinical reasoning rooted in probabilistic inference and democratic consensus—paralleling the uncertainty navigation required in AI-driven decision-making [59]. These perspectives suggest that ethical, transparent integration of AI will require interdisciplinary collaboration between clinicians, data scientists, ethicists, and policymakers to maximise patient benefit and public trust.

A summary of the key findings related to AI driven precise and tailored interventions and areas for future research can be found in

Table 4. AI Applications in Personalised Therapy, Surgery, and Rehabilitation: Summarises AI-driven advances in pharmacogenomics, robotic-assisted transplantation, and personalised rehabilitation, with recommendations for optimising model complexity, scaling surgical infrastructure, and implementing ethical governance frameworks.

Key Findings

AI models, particularly those incorporating pharmacogenomics, enable tailored immunosuppressive therapy by analysing genetic variants such as CYP3A5 to optimise tacrolimus dosing. AI-driven robotic systems improve transplant surgical accuracy, reduce complications, and support consistent organ placement. Post-surgery, AI enables personalised rehabilitation by analysing multi-source data to monitor recovery, predict complications, and support behavioural interventions, such as structured walking programmes.

Recommendations

Optimise AI models to capture dynamic gene–environment–treatment interactions in pharmacogenomics. Develop scalable infrastructure for AI-assisted robotic surgery, with a focus on reducing implementation costs, standardising training, and ensuring equitable access across healthcare systems. Design and operationalise ethical governance frameworks to guide the deployment of AI technologies within national transplant systems.

.

6. AI in Hypothermic and Normothermic Machine Perfusion

6.1. Predictive Modelling for Organ Viability

The integration of AI into machine perfusion (MP) technologies is advancing our ability to predict graft viability before transplantation, particularly for marginal organs. Experimental studies by Asong-Fontem et al. in rat liver models demonstrated that pre-ischemic hypothermic oxygenated perfusion (HOPE-PRE) enhances hepatocyte energy reserves, while post-ischemic HOPE (HOPE-END) better mitigates ischemia-induced inflammation and apoptosis [60]. These findings indicate that AI models trained on perfusion-phase data could help personalise perfusion strategies based on graft condition.

This concept is reinforced by Martins et al., who noted that normothermic machine perfusion (NMP) maintains liver metabolic activity under physiological conditions, creating an ideal setting for viability assessment and AI modelling [61]. Foguenne et al. further bridged experimental and clinical perspectives by detailing how both continuous and end-ischemic HOPE and NMP are currently used in human liver transplant programmes, underlining the translational relevance of perfusion data for AI-based outcome prediction [62].

Deep learning models are increasingly capable of processing these complex, real-time perfusion datasets. As highlighted by Tatum et al., modern MP systems are evolving into dynamic platforms capable of diagnostic-level organ assessment when enhanced with AI, enabling integration of biochemical, hemodynamic, and metabolic markers [63]. Such models may eventually underpin standardised, data-driven viability assessments across institutions.

6.2. Real-Time Organ Monitoring and Perfusion Optimisation

AI’s ability to facilitate real-time decision-making is particularly beneficial during the perfusion process. Integrated ML systems can adjust perfusion parameters—such as temperature, oxygenation, and flow rates—in response to minute-to-minute physiological feedback. For example, terahertz spectroscopy coupled with convolutional neural networks has enabled precise apoptosis detection, enabling non-invasive and real-time assessments of tissue viability [64].

Magbagbeola et al. designed a modular and scalable research platform for ex vivo liver NMP, highlighting the potential for AI integration into the architecture of perfusion devices [65]. Such platforms provide the foundation for continuously optimised perfusion protocols, improving graft quality by maintaining homeostasis during preservation.

This approach has also been adapted to surgical contexts involving hypothermic organ protection. In robotic kidney transplantation, intraoperative hypothermia has been shown to mitigate ischemic injury. Navarro et al. evaluated laparoscopic cooling techniques and demonstrated improved viability metrics with regional hypothermia during renal transplantation [66]. These findings support the application of AI-enhanced control systems to dynamically regulate perfusion and temperature during transplantation procedures.

Taylor and Baicu provided a clinical outlook on hypothermic perfusion, arguing that machine perfusion augmented by AI could transition from passive preservation to active organ conditioning, potentially influencing post-transplant outcomes [67].

6.3. AI-Guided Decision-Making in Organ Acceptance

Automated AI systems are also being deployed to inform decision-making regarding marginal organ acceptance. By evaluating donor characteristics, ischemia duration, and perfusion biomarkers, ML models can stratify risk and predict post-transplant outcomes, assisting clinicians in making time-sensitive decisions. Tusch highlighted how sequential decision-making models can improve early-stage triage and organ allocation [68].

In parallel, Parente et al. showed that combining NMP with immune modulation strategies can restore marginal organs, and AI systems could play a pivotal role in monitoring and optimising these interventions through real-time assessment of immunologic and metabolic viability [69]. For islet transplantation, tools like IsletSwipe have already improved the standardisation and reproducibility of graft quality assessments using AI-driven imaging analysis [70].

Wagner et al. extended the utility of AI-guided perfusion systems to thoracic transplantation, illustrating how viability algorithms can support reconditioning protocols, especially in high-risk recipients [71]. Weissenbacher et al. advocated for a broader paradigm shift from organ replacement to reconditioning. They proposed that future AI platforms might not only evaluate organ viability but actively generate and implement reconditioning protocols during perfusion, expanding the criteria for acceptable grafts [72].

Collectively, these advancements highlight how AI is reshaping organ assessment—moving from static evaluation toward continuous, personalised, and potentially therapeutic interventions during the preservation phase.

A summary of the key findings related to AI in hypothermic and normothermic machine perfusion and areas for future research can be found in

Table 5. AI Integration in Machine Perfusion and Organ Viability Assessment: Summarises AI-enhanced prediction, monitoring, and decision-making in machine perfusion, with recommendations for model validation, biomarker integration, and comparative studies on marginal organ acceptance.

Key Findings

Integration of AI into hypothermic and normothermic machine perfusion systems enables accurate, real-time prediction of graft function and viability by analysing complex physiological data, improving assessment, particularly for marginal organs. ML embedded in perfusion devices supports dynamic adjustment of parameters (e.g., temperature, oxygenation), reducing ischemic injury and enhancing graft preservation through continuous, automated control. AI-driven decision-support tools facilitate early risk assessment and organ acceptance by evaluating donor, perfusion, and immunologic factors, with growing potential to guide active reconditioning protocols during machine perfusion.

Recommendations

Develop robust validation strategies to standardise AI-assisted organ viability models across transplant centres and organ types. Design integrative AI frameworks that incorporate immunologic and metabolic biomarkers to support dynamic, intra-perfusion reconditioning protocols for marginal grafts. Conduct comparative outcome studies to evaluate the effectiveness, safety, and efficiency of AI-guided versus clinician-guided decision-making in the acceptance of marginal organs.

.

7. Challenges, Limitations and Ethical Considerations

7.1. Data Privacy and Algorithmic Bias

As AI technologies become more deeply embedded in transplantation and broader healthcare systems, issues related to data privacy and algorithmic fairness become increasingly critical. Large Language Models (LLMs) trained on vast, often heterogeneous datasets—many of which include sensitive patient data like genetic information and clinical histories—raise risks of data misuse and re-identification, even from anonymised inputs [73,74]. These concerns are particularly acute in domains such as evidence synthesis and risk of bias (RoB) assessments, where inaccuracies can translate directly into compromised clinical decisions. For instance, failure to detect trial biases in stroke medication studies has been associated with misguided treatment strategies, resulting in avoidable patient harm or even death [75].

Bias embedded in training datasets can exacerbate health disparities by yielding AI tools that perform poorly on underrepresented populations [76]. Such imbalances, often originating from systemic inequalities in data collection and healthcare access, can be inadvertently perpetuated or amplified by AI. Mitigating these risks demands deliberate efforts to ensure dataset diversity, transparency in data provenance, and adherence to ethical guidelines for equitable model development [77]. Addressing privacy and bias is foundational to the ethical deployment of AI, particularly in high-stakes fields like transplantation, where errors can have irreversible consequences.

7.2. Barriers to Clinical Integration and Regulatory Adoption

Despite the growing promise of AI, integration into clinical practice remains limited due to systemic, regulatory, and institutional barriers. A prime example is the slow adoption of automated knowledge-based planning (KBP) in radiotherapy, which—despite showing equivalence or superiority to human-generated plans in consistency and performance—has been hampered by a lack of standardisation, validation, and institutional confidence [78]. This example mirrors broader resistance in clinical transplantation workflows, where similar AI-driven models face scepticism and delayed uptake.

The absence of robust external validation presents a significant risk, as underlined by Bleeker et al., who caution that premature model implementation can compromise patient safety and erode trust in AI-assisted care [79]. Legal ambiguity surrounding accountability further complicates adoption. As AI begins to shape critical medical decisions, the question of liability—whether borne by developers, clinicians, or healthcare institutions—remains unresolved [73,76]. This uncertainty deters institutions from fully integrating AI into clinical practice and underscores the need for legal and regulatory frameworks that clearly delineate responsibility in AI-assisted medicine.

7.3. Transparency, Interpretability, and the Need for Explainable AI

One of the most pressing ethical challenges in the clinical use of AI is the opacity of many machine learning systems. Complex algorithms—especially deep learning networks—often operate as “black boxes,” offering little insight into how conclusions are derived [74,76]. This lack of transparency undermines trust, impedes auditing, and obstructs clinician understanding, which is especially problematic in transplant medicine, where decisions are often urgent and life-critical.

In contexts like clinical trials or transplant eligibility scoring, the inability to trace or justify algorithmic decisions introduces unacceptable risk. Without explainability, even accurate AI outputs can become ethically and clinically untenable. Modern ethical frameworks and regulatory expectations now emphasise the need for AI systems to prioritise interpretability, auditability, and human oversight [80]. These principles are not only instrumental in maintaining clinical integrity but are also increasingly being embedded in medical AI regulations to align with public expectations and patient rights.

A summary of the key findings related to the challenges, limitations and ethical Considerations of AI and areas for future research can be found in

Table 6. Ethical and Implementation Challenges in AI-Driven Transplantation: Outlines risks of data re-identification, bias, lack of interpretability, and premature clinical deployment, with recommendations for standardised validation, clinician-AI balance, and development of explainable AI frameworks.

Key Findings

Anonymised patient data may still be vulnerable to re-identification, and biassed training datasets can result in skewed healthcare outcomes, particularly affecting underrepresented demographic groups. Premature deployment of AI models without sufficient external validation can undermine trust and patient safety. The lack of interpretability makes it difficult to audit decisions and ensure transparency, highlighting the need for AI models in healthcare to be explainable, accountable, and subject to human oversight to ensure patient safety and compliance with ethical standards.

Recommendations

Research should focus on creating standardised validation frameworks for AI models across diverse clinical settings and patient demographics to ensure reliable and equitable outcomes. Further studies are needed to explore the optimal balance between AI assistance and clinician decision-making, particularly in high-stake clinical decisions. Research should prioritise the development of XAI frameworks for critical healthcare applications, focusing on transparency, interpretability, and the ability to audit AI decision-making processes.

.

8. Future Perspectives and Research Directions

8.1. Enhancing Real-Time Monitoring in Transplant Care

The future of transplant care is closely tied to the development of AI systems capable of real-time decision-making. As surgical complexity and post-operative management demands increase, AI must evolve to handle high-frequency, heterogeneous clinical data streams. Sáez et al. emphasise the critical importance of robust data quality protocols—such as imputation for missing values and data augmentation techniques—which are essential in building reliable real-time monitoring systems for transplantation [81]. These methods address common clinical data challenges, including temporal gaps and inconsistent documentation, especially prevalent in multi-organ transplant scenarios.

In parallel, advancements in human-AI interaction will be essential. Dialogue-based AI systems that can anticipate clinician needs and adapt dynamically to partial or uncertain data inputs are likely to become integral to transplant care. These systems, when designed with a strong user-centric focus and supported by continual learning mechanisms, could evolve in sync with patient conditions—enabling timely, context-aware recommendations during perioperative and postoperative monitoring [81].

8.2. Integrating Multi-Omics Data for Precision Transplant Medicine

AI’s role in transplant monitoring is poised to expand significantly through the integration of multi-omics data—including genomics, proteomics, metabolomics, and transcriptomics. These molecular layers offer detailed insight into patient-specific immune responses, drug metabolism, and risk of complications. However, translating this complexity into actionable intelligence presents substantial challenges. AI approaches such as unsupervised learning are well-suited to uncovering hidden patterns in these datasets, enabling the stratification of patient subgroups and biomarker discovery for personalised treatment planning [81].

To ensure clinical applicability, AI models must also be able to quantify uncertainty in predictions, particularly when extrapolating to unseen patient populations. Strategies such as uncertainty-aware modelling, transportability assessments, and real-world validation will be essential to building generalisable tools. Moreover, federated learning offers a privacy-preserving solution for model training across institutions, helping address regulatory and data-sharing barriers while maximising dataset diversity [81]. These approaches together can facilitate the safe and effective use of multi-omics-informed AI in clinical transplant practice.

8.3. AI and Machine Perfusion: Towards Intelligent Organ Preservation

Machine perfusion (MP) technologies—whether hypothermic or normothermic—are revolutionising organ preservation, and their synergy with AI stands to further improve graft outcomes. Real-time data from perfusion systems, such as flow rates, oxygenation, and biochemical markers, can be continuously analysed by AI algorithms to assess organ viability and suggest dynamic adjustments to perfusion parameters. This feedback loop can enhance organ preservation protocols and reduce variability in outcomes.

Looking beyond real-time analysis, AI-enabled perfusion systems could simulate historical outcomes to develop predictive models for tailoring preservation strategies to individual grafts. Januszewski and Jain suggest that large-scale AI foundation models could eventually identify subtle perfusion trends across diverse transplant cases, contributing to better donor organ utilisation and outcome forecasting [82]. Complementary to this, model compression techniques—such as knowledge distillation—can help transfer capabilities from large, complex AI models to smaller, task-specific models, improving computational efficiency and enabling integration into perfusion hardware [82].

As AI-perfusion synergy matures, the next-generation platforms will not only evaluate but also suggest optimised preservation regimens, expanding the clinical use of machine perfusion for marginal organs and increasing transplantation success rates.

A summary of future perspectives and research directions can be found in

Table 7. Future Research Priorities for AI in Transplant Monitoring: Highlights the need for scalable, multi-omics-integrated, and efficient AI systems capable of real-time monitoring, handling missing data, and addressing ethical concerns via federated and unsupervised learning approaches.

Key Findings

AI-driven transplant monitoring systems must evolve to handle large, dynamic datasets efficiently to support proactive and personalised clinical decision-making. Future research should explore AI models capable of integrating multi-omics data (genomics, proteomics, metabolomics, and transcriptomics) to improve transplant monitoring. Research should focus on improving computational efficiency and the scalability of AI models for real-time applications, ultimately enhancing transplant outcomes through more tailored protocols.

Recommendations

Research should focus on developing and validating AI models that can efficiently handle missing or imprecise clinical data through techniques such as data imputation and augmentation. Special emphasis should be placed on developing unsupervised learning techniques for identifying biological subgroups and handling data uncertainty while addressing ethical and privacy concerns through federated learning approaches. AI-powered prediction models that incorporate organ-specific parameters and computational techniques like knowledge distillation, ensuring that these systems are practical for widespread clinical adoption and enhance long-term transplant outcomes, should be explored.

.

9. Conclusions

Artificial Intelligence (AI) is transforming transplantation across the care continuum—enhancing donor–recipient matching, immune risk stratification, immunosuppressive management, surgical precision, and post-operative care. By integrating multi-omics data, electronic health records (EHRs), and wearable devices, AI-driven platforms enable early rejection detection, personalised immunosuppression, and risk-adapted follow-up schedules. AI models have outperformed traditional diagnostic tools in identifying rejection, graft dysfunction, and pharmacogenomic responses, particularly in tailoring tacrolimus dosing. Robotic systems and AI-assisted rehabilitation strategies further extend precision care to surgical and recovery phases, while AI-integrated decision-support systems (DSS) improve candidate selection by augmenting conventional scoring systems such as MELD and KDRI.

Despite these advances, several challenges must be addressed before broad clinical adoption. Data privacy, algorithmic bias, lack of model transparency, and legal concerns around clinician liability continue to impede implementation. Many AI systems remain “black boxes,” necessitating the development of explainable AI (XAI) to ensure clinical trust and regulatory compliance. Integrating federated learning and multi-omics analysis can support more robust and generalisable models while preserving data security. Finally, the synergy between AI and machine perfusion offers new avenues for dynamic organ assessment, especially for marginal grafts. Continued interdisciplinary collaboration, ethical oversight, and real-world validation are essential to fully realise the potential of AI to personalise transplant care and improve long-term outcomes.