Artificial intelligence and machine learning in transplantation surgery care pathway

Abstract

Artificial intelligence (AI) and machine learning (ML) are increasingly applied across the transplantation pathway, offering advances in preoperative planning, perioperative management, and postoperative recovery. In preoperative care, deep learning algorithms improve anatomical assessment, volumetry, and graft weight estimation, while ML-based functional status evaluation and urgency scoring refine candidate selection. Predictive models incorporating metabolic and physiological data further support surgical eligibility and targeted prehabilitation strategies. Perioperatively, ML models outperform conventional approaches in predicting massive transfusion, intraoperative haemorrhage, and acute kidney injury, with explainable outputs enhancing interpretability and clinical trust. Robotic and AI-assisted surgical platforms demonstrate functional equivalence or superiority to conventional methods, reducing intraoperative complications and accelerating recovery, particularly in high-risk cohorts. Postoperatively, ML-driven models enable early prediction of sepsis, pneumonia, and graft dysfunction, while longitudinal markers such as the recipient-to-donor estimated glomerular filtration rate ratio and novel imaging or biomarker-based approaches inform long-term graft monitoring. Optimised perioperative strategies, including analgesic regimens and fluid management, further enhance donor recovery and rehabilitation outcomes. Cross-cutting innovations include imaging-based AI applications such as hyperspectral imaging for real-time graft viability assessment and deep learning for automated histopathological evaluation, which improve diagnostic speed, accuracy, and reproducibility. Multimodal models integrating electronic health records, intraoperative signals, ultrasound, and histology provide dynamic, system-wide insights into graft function and rejection risk, bridging diagnostic, prognostic, and therapeutic decision-making. AI and ML thus hold substantial potential to personalise transplant care and improve outcomes. Their translation into practice, however, requires rigorous validation, dataset diversity, and strong ethical and regulatory governance.

Key Words: Artificial intelligence; Machine learning; Pre-operative planning; Peri-operative care; Rehabilitation

Core Tip: Artificial intelligence (AI) and machine learning (ML) are transforming transplantation by enhancing precision across preoperative, perioperative, and postoperative phases. Deep learning improves anatomical assessment, graft evaluation, and candidate selection, while ML-based models predict intraoperative complications and postoperative risks such as sepsis, graft dysfunction, and renal injury. AI-assisted surgical platforms and multimodal predictive systems integrating imaging, histopathology, and electronic health records further personalise decision-making, optimise recovery, and refine long-term graft monitoring. Successful clinical translation, however, hinges on rigorous validation, diverse datasets, and robust ethical and regulatory oversight.

INTRODUCTION

The integration of artificial intelligence (AI) and machine learning (ML) into transplantation surgery is rapidly reshaping the clinical landscape, from prehabilitation and surgical planning to intraoperative guidance, postoperative monitoring, and long-term outcome prediction. Transplantation remains one of the most complex and resource-intensive domains of modern medicine, where timely and accurate decision-making directly impacts graft viability, patient survival, and health system efficiency. AI and ML are uniquely positioned to augment these processes by enabling real-time data interpretation, predictive analytics, and personalised risk stratification.

Advances in preoperative imaging, such as deep learning-assisted 3D segmentation and volumetric analysis, have improved the precision of anatomical planning in living donor liver transplantation (LDLT), reducing variability and operative risk. Simultaneously, ML-based risk stratification models have shown superior accuracy in predicting perioperative complications compared to traditional clinician judgment, enabling more refined surgical eligibility and resource allocation.

Perioperative and intraoperative AI applications are increasingly being used to forecast complications such as massive transfusion, acute kidney injury (AKI), and coagulopathy using dynamic clinical and physiological data. Robotic and ML-enhanced surgical platforms further support precision and minimally invasive techniques, especially in high-risk populations such as paediatric or obese transplant recipients.

Postoperatively, AI models demonstrate robust performance in forecasting complications such as sepsis, pneumonia, and delayed graft function (DGF) (defined as the requirement for dialysis within the first week post-transplant), guiding early interventions and improving clinical outcomes. Longitudinal tracking models based on early graft function or donor-recipient glomerular filtration rate (GFR) ratios enable refined prognostication of renal recovery and graft survival. Additionally, AI-enhanced imaging modalities and histopathology tools offer unprecedented accuracy and speed in assessing organ viability and pathology intraoperatively and postoperatively.

This narrative review synthesises recent advancements in AI and ML across the transplantation care pathway, highlighting evidence-based applications in preoperative planning, intraoperative risk prediction, postoperative monitoring, functional recovery, and long-term outcome modelling (Figure 1). It also explores the emerging role of cross-disciplinary and multiorgan AI innovations, setting the foundation for future personalised, data-driven transplant care (Table 1).

Figure 1 The transplantation surgery care pathway and mapped artificial intelligence/machine learning applications. The figure summarises the principal artificial intelligence (AI) and machine learning (ML) applications discussed in this review, mapped onto the three phases of the pathway (preoperative, perioperative/intraoperative, postoperative), with cross-cutting requirements for safe clinical translation displayed below. AKI: Acute kidney injury; ANN: Artificial neural network; CKD: Chronic kidney disease; DGF: Delayed graft function; DL: Deep learning; DRHF: Dynamic ratio of hepatic function; eGFR: Estimated glomerular filtration rate; GBM: Gradient boosting machine; GRWR: Graft-to-recipient weight ratio; GV: Graft volume; SLV: Standard liver volume; HAO: Hepatic artery occlusion; KPS: Karnofsky performance status; LDLT: Living donor liver transplantation; ML: Machine learning; MRCP: Magnetic resonance cholangiopancreatography; RAKT: Robot-assisted kidney transplantation; RNN: Recurrent neural network; SHAP: Shapley additive explanations; NMR: Nuclear magnetic resonance.

Table 1 Summary of artificial intelligence and machine learning studies mapped to the transplantation surgery care pathway, organised by operative phase.

Ref.	Application/focus	Key findings	Limitations & future directions
Preoperative imaging & planning
Oh et al[1]	DL model for automated 3D biliary segmentation (MRCP) in LDLT donors	DSC = 0.80 ± 0.20; reduced manual workload; inferred missing duct segments improving clarity for CHD/hilum	Trained on homogeneous cohort (young, ideal BMI); needs validation in diverse populations and clinical integration
Park et al[2]	DL-assisted CT volumetry for graft weight estimation	CCC = 0.834 vs actual weights; mean processing time 18 minutes; nearly perfect inter-reader agreement (CCC = 0.998)	Minor underestimation in male/high-BMI donors; focused only on right-lobe grafts; external validation required
Machry et al[3]	Review of AI volumetric tools in LDLT	AI platforms (e.g., “Dr. Liver”) reduced volumetry time (38 minutes → 7 minutes); improved accuracy (error 3.1% vs > 10% manual)	Limited evidence for left lobe/paediatric grafts; requires broader adoption and workflow integration
Risk stratification & functional status
Brennan et al[4]	MySurgeryRisk ML algorithm vs physician judgment	Improved AUC; reduced under/overestimation of complications; transparent UI improved trust	Single-centre study; prospective validation ongoing
Chu et al[5]; Thongprayoon et al[6]	ML analysis of KPS in > 224000 KT recipients	Post-KT functional improvement approximately 0.89%/year; low KPS strongly predicted mortality and graft loss	KPS predictive validity (C-statistic approximately 0.60) < frailty/performance measures (approximately 0.75); annual pre-KT screening advised
	Unsupervised ML clustering of low KPS (≤ 40%)	Cluster 2 (younger, more living donors, fewer HLA mismatches) showed superior 5-year outcomes vs cluster 1	Highlights systemic inequities; challenges exclusion of low-functioning candidates; needs socio-clinical integration
Lyden et al[7]	Statistical best practices for ML urgency scoring	Advocates time-dependent covariates, cause-specific hazard models, proper handling of waitlist mortality	Requires rigorous adherence to statistical standards to avoid immortal time bias and misclassification
Prehabilitation & surgical eligibility
Krivov et al[8]	¹H NMR metabolomic tracking for kidney transplant recovery	Differentiated primary vs delayed graft function by POD 2; diffusion framework robust against overfitting	Limited resolution for acute rejection; requires enhanced temporal resolution
Jung et al[9]	ML model for 90-day graft failure in paediatric LT	AUROC = 0.898; key predictors: Encephalopathy, sodium, bilirubin, vascular thrombosis; enabled perioperative stratification	Retrospective, single-centre; requires multicentre validation
Law and Kow[10]	Risk factors for SFSS in LDLT	Donor age > 45, macrosteatosis > 10%, high portal venous pressure = higher risk; GRWR ≥ 0.8%, GV/SLV ≥ 40% protective	BMI not predictive unless with steatosis; AI volumetry needed for accurate planning
Prediction of intraoperative complications
Chen et al[11]	CatBoost ML model for massive transfusion in LT	Outperformed logistic regression/XGBoost; key predictors: Hb, Hct, fibrinogen, platelets	Lower specificity in prospective data; excluded intraoperative variables; multicentre validation needed
Eyth et al[12]	Manually calculable risk score for RBC transfusions	Incorporated Hb ≤ 7.5 g/dL, ASA, urgency, hypoalbuminaemia, thrombocytopenia; aids planning & resource allocation	Not AI-driven; performance may vary across surgeries; requires validation in broader LT cohorts
Park et al[13]	ML risk score for intraoperative haemorrhage	Predictors: MELD, aPTT, creatinine, MAP, pulse pressure, temperature; PT/INR not predictive	Retrospective study; prospective validation required
Zhang et al[14]	GBM + SHAP for intraoperative AKI risk	Strong predictors: Urine output (approximately 2.2 mL/kg/hour), anaesthesia time, platelet count; SHAP confirmed validity	Single-centre; external validation needed
Liu et al[15]	GBM + SHAP for AKI risk in DCD grafts	Predictors: Indirect bilirubin, urine output, anaesthesia time, platelets, steatosis; superior to other ML models	No external validation; lacked AKI staging
Lee et al[16]	GBM model for AKI in LT	Key predictors: Cold ischaemic time, SvO2; online calculator developed	Prospective testing pending
Bredt et al[17]	ANN vs logistic regression for post-LT AKI	ANN outperformed LR; predictors: MELD, marginal grafts, hypotension, transfusion, perfusion	Retrospective, single-centre; limited interpretability/generalisation
Robotic & ML-enhanced surgical techniques
Martucci et al[18]	Da Vinci robotic hepatectomy in LT	Enabled precision, 3D vision, EndoWrist dexterity; moderate blood loss; feasible anaesthetic optimisation	High cost, training requirements, physiological challenges (e.g., pneumoperitoneum, hypothermia)
Territo et al[19]	RAKT vs OKT	Lower blood loss, pain, wound complications; equivalent renal outcomes despite longer ischaemia times	Higher operative time; requires technical expertise
Slagter et al[20]	RAKT in obese recipients	Lower SSI, fewer lymphoceles, shorter hospital stay; learning curve approximately 21-35 cases	High cost; contraindicated in calcified iliac arteries; limited to selected patients
Karadag et al[21]	RAKT vs OKT	Less bleeding, fewer complications, shorter hospital stays; longer ischaemia/anastomosis times	Requires experienced centres; standardisation needed
Khajeh et al[22]	RADN vs LDN	Better outcomes in operation time, complications, and recovery after learning curve; slightly more blood loss	Surgeon experience crucial; need standardised definition of ‘experience’
Giulianotti et al[23]	Robotic surgery in obese transplant patients	Reduced incision size, trauma, wound complications; expanded surgical feasibility	High cost, warm ischaemia risks; requires long-term data
Mulloy et al[24]	Robotic transplant nephrectomy	Lower blood loss, infections, hospital stays; minimally invasive alternative	No FDA-approved robotic staplers; higher costs, longer operative times
Predicting postoperative complications & graft function
Chen et al[27]	ML (random forest) for sepsis after LT	Identified 8 risk factors; RF outperformed SOFA (AUC 0.731); online calculator developed	Retrospective, single-centre; moderate sensitivity; excluded pre-operative sepsis
Luo et al[28]	Random forest for pneumonia after KT	AUROC 0.91; predictors: Pulmonary lesions, reoperation, rATG dose, albumin, Ig, DGF; guided escalation protocol	Small, single-centre; retrospective; class imbalance
Tan et al[29]	Nomogram for paediatric LDLT graft recovery	DRHF + 4 predictors (donor liver insufficiency, ischaemia time, creatinine, bilirubin); AUROC 0.898	Single-centre, small sample; requires multicentre validation
Longitudinal monitoring of recovery
Wan et al[30]	Recipient-to-donor eGFR ratio	RFR ≥ 1 Linked to 45% lower 10-year graft failure; optimal 1-2; outside range = inadequate recovery/hyperfiltration	Single-centre; small sample; multicentre validation needed
Campbell et al[31]	Predictors of renal recovery post-OLT	Duration of renal dysfunction > 3.6 weeks predicted poor 12-month renal function (AUC 0.71); CKLT showed better recovery	Creatinine-based GFR limits; retrospective
Levitsky et al[32]	Native vs graft GFR in SLK	Radionuclide scans improved discrimination; abnormal renal imaging predicted poor recovery; exploratory biomarkers proposed	UNOS criteria insufficient; biomarkers unvalidated; small sample
Functionality & rehabilitation
Park et al[33]	Analgesic impact (ITMB) in living kidney donors	ITMB reduced delayed kidney recovery by 74.3%; protective via haemodynamic stabilisation; fluid balance also critical	Mechanism indirect; limited validation; demographic disparities noted
Predictive models of outcomes & survival
Molinari et al[34]	Mortality risk model in cadaveric LT	C-statistic up to 0.95 for 90-day mortality; excludes sex/diagnosis to reduce bias	Retrospective; lacks frailty/nutrition data; needs validation
Sliwinski et al[35]	Prehab App risk calculator	Simple metrics predicted 90-day survival; < 5 minutes use; supports personalised prehabilitation strategies	Single-centre; excluded cardiac/orthopaedic cases; AI-enhancements planned
Imaging-based AI applications
Felli et al[36]	AI-enhanced hyperspectral imaging for liver viability during HAO	Accurate tissue oxygenation mapping; strong correlation with lactates & Suzuki score; CNN: Sensitivity 0.993, specificity 0.997	Shallow penetration (3-5 mm); no real-time video; limited ischaemia discrimination; future HYPER-reality integration proposed
Marsh et al[37]	Deep learning for glomerulosclerosis quantification in kidney biopsies	Fully convolutional model reduced processing time 22-fold; improved pixel-wise accuracy; reduced inter-observer variability	Misclassification near vessels/cysts; requires larger datasets and refined post-processing
Multiorgan & multidisciplinary applications
Rashidi and Bihorac[38]	RNNs for AKI prediction using EHR and intraoperative data	Predicted AKI up to 48 hours pre-onset; dynamic recalibration with evolving physiology; integrated intra-operative signals for continuous risk	Needs broader validation; integration into perioperative workflows still limited
	CNNs for CKD stage classification via ultrasound	Non-invasive eGFR and CKD staging; reduced reliance on serum creatinine; radiation-free monitoring	Requires validation across diverse populations; long-term performance unknown
	CNNs for renal histopathology segmentation	Automated quantification of glomeruli, tubules, interstitium, vasculature; supports rejection/fibrosis grading	Risk of variability with rare morphologies; dataset standardisation required

AKI: Acute kidney injury; ANN: Artificial neural network; AUC: Area under the curve; AUROC: Area under the receiver-operating characteristic curve; CCC: Concordance correlation coefficient; CKLT: Combined kidney-liver transplant; DCD: Donation after circulatory death; DGF: Delayed graft function; DSC: Dice similarity coefficient; GBM: Gradient boosting machine; GFR: Glomerular filtration rate; GRWR: Graft-to-recipient weight ratio; GV: Graft volume; SLV: Standard liver volume; HAO: Hepatic artery occlusion; KPS: Karnofsky Performance Score; KT: Kidney transplantation; LDLT: Living-donor liver transplantation; LT: Liver transplantation; ML: Machine learning; NMR: Nuclear magnetic resonance; OKT: Open kidney transplantation; POD: Postoperative day; RAKT: Robot-assisted kidney transplantation; RADN: Robot-assisted donor nephrectomy; RFR: Recipient-to-donor functional ratio; RF: Random forest; RNN: Recurrent neural network; SFSS: Small-for-size syndrome; SHAP: Shapley additive explanations; SLK: Simultaneous liver-kidney transplant; SSI: Surgical site infection; MRCP: Magnetic resonance cholangiopancreatography; DL: Deep learning; CHD: Common hepatic duct; BMI: Body mass index; CT: Computed tomography; AI: Artificial intelligence; UI: User interface; XGBoost: Extreme gradient boosting; Hb: Haemoglobin; ASA: American Society of Anesthesiologists; RBC: Red blood cells; MELD: Model for end-stage liver disease; aPTT: Activated partial thromboplastin time; MAP: Mean arterial pressure; PT: Prothrombin time; INR: International normalized ratio; FDA: Food and Drug Administration; SOFA: Sequential organ failure assessment; rATG: Rabbit anti-thymocyte globulin; DRHF: Dynamic ratio of hepatic function; OLT: Orthotopic liver transplantation; ITMB: Intrathecal morphine block; CNN: Convolutional neural network; EHR: Electronic health record.

LITERATURE SEARCH

A structured but non-systematic literature search was conducted in PubMed to identify peer-reviewed articles published between 2005 and 2025. The search strategy combined Medical Subject Headings and free-text keywords related to AI, ML, and transplantation. Search terms were tailored to four thematic domains and their subcategories, as detailed below. As this is a narrative review, a formal the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) flow was not undertaken; we explicitly do not claim a comprehensive systematic search.

AI and ML in prehabilitation and preoperative planning

(1) Subdomains: Preoperative Imaging and planning; risk stratification and functional status evaluation; prehabilitation and surgical eligibility; and (2) Search string: (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“prehabilitation” OR “preoperative planning” OR “imaging” OR “volumetry” OR “risk stratification” OR “functional status” OR “surgical eligibility”) AND (“transplantation” OR “transplant surgery”).

AI and ML in peri-operative care and intraoperative risk prediction

(1) Subdomains: Prediction of intraoperative complications; robotic and ML-enhanced surgical techniques; and (2) Search string: (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“perioperative” OR “intraoperative” OR “risk prediction” OR “complications” OR “robotic surgery”) AND (“transplantation” OR “liver transplant” OR “kidney transplant”).

AI and ML in postoperative monitoring and early recovery

(1) Subdomains: Predicting postoperative complications and graft function; longitudinal monitoring of recovery; functionality and rehabilitation; predictive models of outcomes and survival; and (2) Search string: (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“postoperative” OR “monitoring” OR “recovery” OR “rehabilitation” OR “graft function” OR “survival prediction”) AND (“transplantation”).

Cross-cutting methodologies and advances

(1) Subdomains: Imaging-based AI applications; multiorgan and multidisciplinary applications; and (2) Search string: (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“hyperspectral imaging” OR “digital pathology” OR “histopathology” OR “multiorgan” OR “multidisciplinary” OR “cross-cutting”) AND (“transplantation”).

Articles were included if they: (1) Reported original research, systematic reviews, or meta-analyses applying AI/ML in transplantation or perioperative surgical care; (2) Focused on one or more of the specified thematic domains; and (3) Were written in English and involved human subjects or clinically relevant translational models.

Exclusion criteria comprised studies unrelated to transplantation, those without explicit AI or ML methodology, and non peer-reviewed sources such as editorials, conference abstracts, and preprints. Records were assembled into a single library and de-duplicated by digital object identifier and PubMed identifier. Two stages of screening were applied: Title and abstract review against the inclusion and exclusion criteria, followed by full-text appraisal of potentially eligible studies. Reference lists of included articles were hand-searched to identify further relevant publications. Studies were selected to reflect methodological and thematic breadth across the transplantation pathway rather than to exhaustively enumerate all available evidence; we therefore do not report screening counts in PRISMA format. This narrative selection inevitably carries a risk of selection bias, which we acknowledge as a limitation of the review.

AI AND ML IN PREHABILITATION AND PREOPERATIVE PLANNING

Preoperative imaging and planning

Accurate preoperative imaging and volumetric assessment are critical for optimising donor and recipient outcomes in LDLT, particularly to mitigate risks such as small-for-size syndrome (SFSS). Recent advances in AI and deep learning algorithms (DLA) have enhanced the precision and efficiency of anatomical evaluation, volumetry, and 3D modelling, supporting both surgical planning and intraoperative decision-making.

Oh et al[1] demonstrated the use of a deep learning model for automated 3D segmentation of biliary structures on magnetic resonance cholangiopancreatography (MRCP) in living liver donors. The model achieved a mean dice similarity coefficient (DSC) of 0.80 ± 0.20, reflecting good segmentation performance overall but with limitations in intrahepatic ducts due to their small size and discontinuity on MRCP images. Accurate bile duct segmentation is vital in LDLT because complex biliary anatomy and small anastomotic sites are associated with high complication rates. The automated model reduced manual workload and enhanced anatomical clarity, supporting surgeons of varying experience levels. Notably, the model could infer missing duct segments by learning implicit anatomical relationships from expert-curated data, sometimes outperforming raw MRCP images in visualising key structures such as the common hepatic duct (CHD) and hilum. However, the model was trained on a homogeneous cohort [young donors with ideal body mass index (BMI)], limiting its generalisability. Future work should expand datasets to diverse populations and validate integration into clinical workflows[1].

Park et al[2] validated a deep learning-assisted computed tomography (CT) volumetry method for rapid and accurate graft weight estimation. The algorithm required no manual correction in 70% of cases and minor adjustments in 30%, with a mean processing time of 1.8 minutes. The study developed a volume-to-weight conversion formula based on 207 donors, demonstrating good agreement with actual graft weights [concordance correlation coefficient (CCC) = 0.834] despite a ± 17.1% error margin, which still outperformed previous approaches. Minor underestimation biases were observed in male donors (-2.6%) and those with higher BMI (-4.2%), reflecting demographic differences between training and validation groups. Inter-reader agreement on volumetry was nearly perfect (CCC = 0.998), suggesting DLA can standardise measurements and reduce variability across operator experience. This study focused on right-lobe grafts in a relatively homogeneous population, and external validation across broader donor demographics and graft types remains necessary[2].

Machry et al[3] reviewed AI integration in liver volumetric analysis, emphasising its role in preoperative anatomical assessment. AI applications in liver surgery have predominantly targeted adult right-lobe donor grafts, with less data on left lobe or left lateral segment grafts. In paediatric LDLT recipients weighing less than 10 kg, large-for-size syndrome is a more common concern than SFSS, requiring intraoperative graft downsizing. By contrast, adult LDLT demands precise preoperative volumetry to prevent SFSS. Studies show that surgeons using AI-powered volumetric software can achieve graft volume estimations comparable to those of radiologists (correlation coefficient r > 0.80), enabling more direct surgeon engagement in anatomical planning and resection strategy. The “Dr. Liver” platform, developed in South Korea for surgeon interaction, demonstrated superior accuracy (r = 0.98) and significantly lower absolute volumetric error (3.1% ± 2.8%) than manual volumetry, which exhibited a clinically significant error exceeding 10% in 46% of cases. Furthermore, Dr. Liver markedly reduced volumetry time from approximately 38 minutes to 7 minutes, highlighting AI’s potential to streamline workflows while improving accuracy in LDLT planning[3].

Advances in AI-driven imaging and volumetric analysis have clearly improved the precision, efficiency, and reproducibility of preoperative planning in LDLT, yet several limitations temper their current clinical impact. Automated segmentation and volumetry tools consistently reduce manual workload and inter-operator variability, with accuracy levels that rival or exceed radiologists, thereby enabling more consistent surgical planning even among less experienced clinicians. Importantly, by enhancing anatomical delineation and graft volume estimation, these models directly address key complications such as biliary injury and SFSS, underscoring their clinical relevance. However, most algorithms remain trained and validated in narrow, homogeneous cohorts, predominantly young, low-BMI, right-lobe donor populations, raising concerns about their generalisability to broader demographics, left lobe grafts, and paediatric recipients where anatomical and physiological challenges differ substantially. Furthermore, while platforms such as “Dr. Liver” demonstrate remarkable gains in speed and accuracy, evidence of integration into real-world surgical workflows remains limited. Thus, while the collective findings establish AI as a transformative adjunct for anatomical assessment, broader dataset diversity, multicentre validation, and workflow integration will be essential for these tools to move beyond proof-of-concept and meaningfully reshape preoperative planning in transplantation.

RISK STRATIFICATION AND FUNCTIONAL STATUS EVALUATION

Effective risk stratification and functional status evaluation are crucial for optimising transplant outcomes and resource allocation. Emerging ML tools and nuanced analyses of functional metrics have enhanced prediction accuracy and clinical decision support in transplantation.

Brennan et al[4] compared physician clinical judgment against an ML-driven algorithm, MySurgeryRisk, for predicting postoperative complications using electronic health record (EHR) data. The algorithm matched or exceeded physician accuracy, particularly by reducing underestimation of true complications and overestimation of non-events. Interaction with the tool significantly modified physicians’ risk assessments, improving classification metrics such as area under the curve (AUC). The algorithm’s transparent user interface, which highlighted key predictive variables, fostered physician trust and interpretability. Interestingly, physicians with higher cognitive reflection scores were more likely to adjust their risk evaluations, indicating that individual decision styles influence the adoption of algorithmic support. Although tested in a small, single-centre cohort, MySurgeryRisk was found user-friendly and is currently integrated into real-time clinical workflows in ongoing prospective studies[4].

Functional status evaluation in kidney transplant (KT) recipients has been advanced by ML approaches tracking longitudinal trends and patient clustering. Chu et al[5] and Thongprayoon et al[6] analysed over 224000 KT recipients, revealing that 39.9% experienced pretransplant declines in Karnofsky Performance Status (KPS) (a clinician-rated 0-100 scale of functional capacity), a widely used but subjective functional metric. Post-transplant, functional status improved at an average rate of approximately 0.89% per year, suggesting meaningful recovery. Notably, improvements were more pronounced in younger, female, non-Black, diabetic, and deceased-donor recipients, reflecting potential biological, social, or systemic disparities. Lower KPS at transplant admission and particularly lower post-KT KPS strongly predicted increased mortality and all-cause graft loss, with post-KT functional changes exhibiting stronger prognostic value. However, KPS’s predictive validity (c-statistic approximately 0.60) was inferior to objective frailty or physical performance measures (C-statistics approximately 0.75), questioning its utility as a standalone risk tool. Given functional status’s dynamic nature, these findings support annual pre-KT functional screening and prehabilitation interventions to mitigate risks in vulnerable populations[5,6].

Using unsupervised ML, these studies identified two phenotypically distinct clusters among recipients with low KPS (≤ 40%). Despite similar baseline functional impairment, cluster 2, characterised by younger age, predominantly white ethnicity, higher living donor transplant rates, and fewer HLA mismatches, demonstrated significantly superior five-year graft and patient survival compared to cluster 1. This difference was associated with higher-quality transplants (shorter cold ischaemia times, more living donors) and suggests that access to better donor organs and supportive resources can attenuate the negative impact of poor functional status. These findings challenge the exclusion of functionally impaired candidates from transplantation, emphasising the importance of integrating functional status with broader clinical and socioeconomic context during risk assessment[5,6].

Lyden et al[7] outlined statistical best practices for ML-based urgency scoring in transplantation, focusing on the importance of aligning modelling approaches with research objectives. When analysing medical urgency, adjustment models should typically include only urgency status, reflecting the existing allocation system without conflating other clinical predictors. Accurate mortality estimates require accounting for deaths during inactive waitlist periods by carrying forward urgency and exception statuses, avoiding bias from censoring. Because candidate status is dynamic, analyses must employ time-dependent covariates to properly model changing urgency, as static covariates introduce immortal time bias. For comparing urgency groups, cause-specific hazard models are preferred over sub-distribution Fine-Gray models, which confound transplant access with urgency status. Finally, outcome selection (pretransplant death, post-delisting death, or composite) should align with study aims, be justified, and sensitivity-tested, supported by comprehensive post-removal mortality data to ensure validity[7].

AI and ML approaches to risk stratification and functional status evaluation offer clear advantages over traditional clinician judgment, yet their clinical maturity remains uneven. Tools such as MySurgeryRisk demonstrate that algorithm-driven prediction can match or exceed expert clinical assessment, particularly for complex outcomes such as AKI, sepsis, and intensive care escalation. Equally important is the recognition that subjective metrics such as KPS, while widely used, capture functional status only coarsely and are vulnerable to inter-rater variability. ML-based clustering has demonstrated that recipients with apparently identical KPS scores can occupy distinct phenotypic groups with markedly different transplant outcomes, supporting the case for richer phenotyping. Emerging objective frailty and physical-performance measures should be incorporated alongside ML pipelines to strengthen this signal: Examples include the Liver Frailty Index (grip strength, chair stands, balance), short physical performance battery, gait speed, CT-derived skeletal muscle index for sarcopenia, and wearable accelerometer-derived activity profiles. These instruments are reproducible, less subjective than clinician-rated scales, and amenable to longitudinal capture both before and after transplantation, providing inputs that an ML model can exploit far more effectively than a single KPS estimate. Together, ML-based clustering and objective frailty measurement offer a route from static eligibility thresholds to dynamic, individualised risk stratification, although prospective validation, integration with electronic records, and demonstration of impact on transplant decisions are still required.

PREHABILITATION AND SURGICAL ELIGIBILITY

Prehabilitation and precise surgical eligibility assessment have emerged as key strategies to optimise transplant outcomes, particularly in complex settings such as paediatric liver transplantation (LT) and marginal graft scenarios. Recent studies emphasise the integration of predictive modelling, metabolic monitoring, and physiological optimisation to improve recipient readiness and mitigate postoperative complications.

Krivov et al[8] introduced a novel approach using a reaction coordinate derived from ¹H NMR spectral dynamics in erythrocyte extracts to track early recovery trajectories following KT. This model effectively differentiated between primary graft function and DGF as early as postoperative day (POD) 2. Unlike principal component analysis, the reaction co-ordinate incorporated temporal information, allowing for probabilistic estimation of clinical outcomes through a diffusion framework on a free energy landscape, an approach well-suited for stochastic biological processes. The model demonstrated robustness against overfitting through leave-one-out cross-validation and maintained performance across different bin sizes and data transformations. Although it struggled to distinguish acute rejection (AR) due to limited temporal resolution and parameter constraints, it highlighted the utility of high-resolution metabolomic tracking, particularly around the creatinine spectral region, as a non-invasive adjunct for early graft monitoring[8].

In paediatric LT, ML approaches have been applied to stratify early graft failure risk and guide perioperative management. Jung et al[9] developed a predictive model for 90-day graft failure using perioperative data through POD 7, employing least absolute shrinkage and selection operator (LASSO) with stability selection and nested cross-validation to manage high dimensionality and collinearity. The resulting logistic regression model achieved strong predictive performance [area under the receiver-operating characteristic curve (AUROC) = 0.898; area under the precision-recall curve = 0.882], and identified key features such as hepatic encephalopathy, intraoperative sodium levels, bilirubin on POD 7, and early vascular thromboses. These variables reflect early graft dysfunction, transfusion impact, and vascular compromise, critical factors in the short post-transplant window. A scoring system derived from these predictors facilitates real-time risk stratification and timely re-transplantation planning. Despite the model’s retrospective and single-centre limitations, it sets a foundation for developing validated, generalisable tools across diverse paediatric cohorts[9].

Risk factors for SFSS, a major concern in LDLT, have also been refined through recent investigations. Law and Kow detailed that older donor age (> 45 years), liver macrosteatosis > 10%, and elevated intraoperative portal venous pressure significantly increase SFSS risk. Importantly, BMI alone is not predictive unless accompanied by hepatic steatosis. Right lobe grafts are now favoured over left lobe grafts due to superior regenerative capacity, fewer complications, and lower SFSS rates, particularly when accompanied by meticulous surgical technique and outflow optimisation. Sarcopenia remains a critical modifiable risk factor, complicating graft size estimation and regeneration post-transplant. Accurate dry weight estimation and AI-enhanced liver volumetry are indispensable for adjusting graft-to-recipient weight ratios (GRWR) and graft volume (GV) to standard liver volume (SLV) thresholds, which should be maintained at GRWR ≥ 0.8% and GV/SLV ≥ 40% to minimise SFSS incidence[10].

Prehabilitation, particularly nutrition optimisation and sarcopenia management, has gained traction as a modifiable preoperative intervention to improve transplant eligibility. These efforts target muscle mass preservation, functional recovery, and reduction of postoperative complications. They are especially valuable in recipients with high model for end-stage liver disease (MELD) scores or acute-on-chronic liver failure, where SFSS risk is amplified. Combining such interventions with dynamic imaging and individualised predictive tools supports a shift from static eligibility criteria toward precision surgical planning.

Prehabilitation and surgical eligibility assessment represent areas where AI and ML applications demonstrate both innovation and clinical necessity, yet the field remains fragmented. High-resolution metabolic monitoring, as illustrated by ¹H NMR-derived recovery trajectories, supports the potential of non-invasive, dynamic biomarkers to complement conventional graft function surveillance, though their inability to reliably discriminate AR highlights a key limitation of early biochemical modelling. Paediatric and complex-graft transplantation deserve particular emphasis. Paediatric LDLT poses challenges that adult-derived models cannot easily extrapolate to: Small recipient size, predominance of left lateral segment grafts, the risk of large-for-size rather than small-for-size physiology in recipients below 10 kg, age-dependent immune responses, and outcomes that must be measured across decades rather than years. AI opportunities in this population include paediatric-specific volumetric atlases for graft downsizing, image-based prediction of biliary and vascular complications (which are disproportionately common in small recipients), longitudinal growth-adjusted graft function modelling, and federated learning across paediatric centres to overcome the small-sample problem inherent to a low-volume specialty. Complex grafts (left lobe, marginal donor, donation after circulatory death, and split or auxiliary grafts) raise analogous issues: Training cohorts are skewed toward adult right-lobe and standard criteria donor anatomy, so model performance in these higher-risk scenarios is at best unproven and at worst misleading. ML-derived scoring systems for early paediatric graft failure such as Jung et al’s model offer a valuable framework for perioperative stratification and proactive re-transplantation planning[9], but their retrospective single-centre origins underscore the need for federated, paediatric-specific multicentre datasets before they can be embedded in clinical pathways. Risk refinement in SFSS has clarified donor- and graft-specific predictors while emphasising the modifiable role of sarcopenia, a factor increasingly recognised as a determinant of transplant success. Here, AI-enhanced volumetry provides a practical bridge between anatomical precision and physiological resilience, enabling graft size optimisation beyond static ratios. Prehabilitation strategies targeting nutrition and sarcopenia extend this paradigm, shifting eligibility from rigid thresholds toward dynamic, modifiable parameters that integrate predictive modelling and individualised care. Together, these findings highlight a transition from descriptive risk factors to actionable, precision-guided eligibility frameworks; validation across diverse populations, including paediatric and complex-graft cohorts, and integration into multidisciplinary protocols remain critical next steps.

AI AND ML IN PERI-OPERATIVE CARE AND INTRAOPERATIVE RISK PREDICTION

Prediction of intraoperative complications

The intraoperative period during LT is associated with substantial risks, including massive transfusion, significant blood loss, and AKI. Recent advances in ML have enabled the development of predictive models to identify high-risk patients and tailor perioperative strategies accordingly.

Chen et al[11] developed the first ML model using the CatBoost algorithm to predict massive transfusion during adult LT. CatBoost outperformed logistic regression and other ML models [e.g., extreme gradient boosting (XGBoost), LightGBM], particularly in handling categorical data and minimising overfitting. Key predictors included haemoglobin, haematocrit, fibrinogen, and platelet count, with haemoglobin contributing most significantly. Despite strong performance, lower specificity in prospective data and the exclusion of intraoperative variables highlight the need for cautious clinical integration and future multicentre validation[11].

Eyth et al[12] introduced a manually calculable risk score for intraoperative packed red blood cell transfusions across a broad range of surgeries, including LT. It incorporated variables such as preoperative anaemia [haemoglobin (Hb) ≤ 7.5 g/dL], American Society of Anesthesiologists (ASA) score, surgical urgency and complexity, hypoalbuminemia, and thrombocytopenia. Actively used in transplant planning, this model facilitates efficient blood preparation and resource allocation, particularly in EHR systems[12].

Park et al[13] developed a risk score using both clinical variables and ML to predict intraoperative haemorrhage. Significant predictors included MELD score, activated partial thromboplastin time, serum creatinine, body temperature, mean arterial pressure, and pulse pressure. Prothrombin time/international normalised ratio was notably not a reliable predictor, aligning with prior evidence. Hypothermia, hypotension, and elevated creatinine were validated contributors to coagulopathy, emphasising the importance of maintaining haemodynamic and thermal stability[13].

Zhang et al[14] developed a gradient boosting machine (GBM) model incorporating 14 variables reflecting pathophysiological mechanisms such as hypovolemia and venous congestion. Intraoperative urine output (approximately 2.2 mL/kg/hour), anaesthesia time, and platelet count were strong predictors. Shapley additive explanations (SHAP) analysis supported the model's clinical validity, suggesting these factors as targets for perioperative renal protection strategies[14].

Liu et al[15] extended this approach to donation after circulatory death (DCD) grafts using GBM and SHAP interpretation. Key predictors included indirect bilirubin, urine output, anaesthesia time, platelet count, and graft steatosis. The model demonstrated superior performance over other ML models, though lacking external validation and AKI staging[15].

Lee et al[16] identified cold ischaemic time and intraoperative SvO₂ as key predictors of AKI using GBM, which outperformed logistic regression and other ML models. An online calculator was developed but awaits prospective testing. The model highlighted oxygen delivery and donor variability as central to AKI risk[16].

Bredt et al[17] found artificial neural networks (ANNs) outperformed logistic regression in predicting post-LT AKI. Major predictors included MELD score, marginal grafts, intraoperative arterial hypotension, massive blood transfusion, and tissue perfusion parameters. While ANNs showed promise, its interpretability and generalisability remain limited by the retrospective, single-centre design[17].

AI and ML approaches for predicting intraoperative complications in transplantation demonstrate clear superiority over conventional statistical models, yet their translation into practice is constrained by methodological and contextual limitations. Across CatBoost, gradient boosting machines, and ANN frameworks, models consistently identify haemodynamic instability, coagulation profiles, transfusion requirements, and tissue perfusion parameters as central predictors, underscoring the shared mechanistic pathways of bleeding and AKI during LT. Explainability tools such as SHAP have begun to bridge the interpretability gap, allowing clinicians to attribute individual predictions to specific variables and validate model outputs against known physiology. To move beyond post-hoc explanation, several groups have begun embedding SHAP-derived feature attributions into real-time decision-support workflows: Continuously streamed intraoperative variables (urine output, mean arterial pressure, SvO₂, temperature, lactate, cumulative transfusion) feed the model at fixed intervals, with per-patient SHAP plots displayed alongside the predicted probability so that the anaesthetic and surgical team can see which physiological signal is driving the rising risk and act on it (for example, by escalating goal-directed fluid therapy or correcting hyperlactataemia). Practical deployment, however, depends on careful integration with the electronic anaesthetic record, transparent thresholds for alerting, and prospective evaluation of whether such alerts change clinician behaviour and patient outcomes rather than simply describing risk. Most published models remain retrospective, single-centre, and often exclude continuous intraoperative variables, limiting real-time utility and generalisability. By contrast, simpler manually calculable scores such as Eyth et al’s transfusion index already support perioperative planning[12], albeit with reduced precision than advanced ML. Together, these findings highlight a maturing evidence base in which ML can enhance intraoperative risk prediction, but clinical adoption will depend on prospective multicentre validation, incorporation of continuous intraoperative signals, and integration into workflow-compatible decision-support systems that balance predictive accuracy with interpretability.

ROBOTIC AND ML-ENHANCED SURGICAL TECHNIQUES

Robotic surgery has become an increasingly significant innovation in perioperative care, offering enhanced precision and reduced invasiveness compared to conventional approaches. Martucci et al[18] extensively described the use of the da Vinci system in robotic surgery, highlighting its components, a surgeon’s console, patient-side robotic arms with EndoWrist instruments offering seven degrees of freedom, and 3D visualisation, which collectively enable superior dexterity and surgical precision beyond traditional methods. However, the high costs of equipment and operation, as well as institutional demands for comprehensive planning, training, and multidisciplinary coordination, remain substantial barriers. Anaesthetic management poses unique challenges due to restricted patient access and physiological alterations such as pneumoperitoneum effects and hypothermia risk. Their report demonstrated the feasibility of fully robotic hepatectomy with optimised anaesthetic protocols, resulting in graft dimensions closely matching preoperative imaging and moderate blood loss[18].

In kidney transplantation, robotic-assisted kidney transplantation (RAKT) has been progressively adopted. Territo et al[19] reported that RAKT, performed intraperitoneally or extraperitoneally depending on patient factors such as obesity, allows graft cooling techniques to mitigate ischaemia-reperfusion injury. Though RAKT operative times are longer or similar to open kidney transplantation (OKT), it results in lower blood loss, less postoperative pain, and fewer wound complications, with comparable or better renal function and low DGF rates. Importantly, extended ischaemia times in RAKT have not correlated with adverse graft outcomes, affirming its functional equivalence to OKT in selected patients[19].

Slagter et al[20] further confirmed RAKT’s advantages, particularly in obese recipients who typically have higher risks of surgical site infections (SSI) and complications. RAKT was associated with significantly lower SSI rates, fewer symptomatic lymphoceles, shorter hospital stays, and reduced pain, without compromising graft or patient survival. Nevertheless, limitations include higher equipment costs and contraindications such as heavily calcified iliac arteries, often present in deceased donor transplants. They emphasised the learning curve for RAKT proficiency, achievable in approximately 21-35 cases by experienced robotic surgeons, and advocated regional hypothermia techniques to counteract longer ischaemia times[20].

Comparative outcome studies support these findings. Karadag et al[21] showed that RAKT has longer ischaemia and anastomosis times than OKT but yields significantly less intraoperative bleeding, shorter hospital stays, and lower postoperative pain. The incidence of postoperative complications was nearly halved in RAKT, with notably fewer wound infections and no lymphoceles observed, highlighting its safety and efficacy in experienced hands. Khajeh et al[22] compared robot-assisted donor nephrectomy (RADN) with laparoscopic donor nephrectomy (LDN), finding that RADN surpassed LDN in operation time, complication rates, and hospital stay once the surgical learning curve was overcome. Although RADN incurred slightly higher intraoperative blood loss, this was not clinically relevant. Their analysis underscored the importance of surgeon experience in mitigating previously reported disadvantages and suggested the need for standardised definitions of “surgical experience” in future research.

Robotic surgery’s utility in complex transplant scenarios was documented by Giulianotti et al[23], who demonstrated that despite the increased perioperative risks in obese patients, robotic techniques reduce incision size and trauma, potentially lowering wound complications and improving recovery. The enhanced dexterity and 3D visualisation allow surgeons with limited laparoscopic experience to safely perform challenging procedures, although high costs and warm ischaemia times remain concerns requiring further optimisation and long-term outcome studies. Similarly, Mulloy et al[24] described robotic transplant nephrectomy as a minimally invasive alternative with superior dexterity and visualisation, resulting in reduced hospital stays, intraoperative blood loss, and wound infections. Challenges include the absence of United States Food and Drug Administration approved robotic vascular staplers and reliance on experienced surgical teams, contributing to longer initial operative times and higher costs, though improvements are anticipated with technological advances and procedural experience.

Robotic platforms are emerging as a transformative adjunct in transplantation, consistently demonstrating reductions in intraoperative bleeding, wound complications, postoperative pain, and length of stay compared with conventional techniques, while maintaining functional equivalence in graft outcomes. The benefits are most pronounced in high-risk populations such as obese recipients, in whom wound morbidity and recovery trajectories are particularly problematic. These advantages appear to outweigh the penalties of longer operative and ischaemia times, provided that surgical teams have surpassed the learning curve (estimated at 20-35 cases for RAKT). Widespread adoption is nonetheless constrained by capital and consumable costs, lack of standardised proficiency definitions, and technical limitations including the absence of robotic vascular staplers. Formal cost-effectiveness data remain limited and heterogeneous; published comparisons of RAKT vs open transplantation suggest higher upfront costs that may be partially offset by shorter inpatient stay and reduced wound-related readmission, although robust incremental cost-effectiveness analyses from prospective comparative cohorts are awaited[25]. Several ongoing and recently completed trials, including multicentre European registries of RAKT in obese recipients and feasibility studies of fully robotic donor hepatectomy, are expected to clarify both clinical and economic value[26]. Perioperative management, particularly anaesthesia and regional cooling strategies, must adapt to robotic contexts to mitigate physiological risks introduced by pneumoperitoneum and prolonged operative times. Overall, robotic surgery offers tangible benefits in safety and recovery, but its long-term value will depend on rigorous health technology assessment, procedural standardisation, and prospective multicentre studies establishing reproducibility and scalability across diverse transplant settings.

AI AND ML IN POSTOPERATIVE MONITORING AND EARLY RECOVERY

Predicting postoperative complications and graft function

ML approaches are increasingly applied to predict critical postoperative complications and graft outcomes following organ transplantation. Recent studies demonstrate the potential of these methods to enable early intervention and personalised management strategies.

Chen et al[27] developed an ML-based model to predict sepsis following LT, a complication affecting 31.9% of patients and associated with heightened perioperative morbidity and mortality, prolonged intensive care unit and hospital stays, and increased healthcare costs. Using LASSO (L1 regularisation) regression, the study identified eight intraoperative and preoperative variables as independent predictors of sepsis, red blood cell transfusion, anaesthesia duration, preoperative total bilirubin, intraoperative blood loss, urine output, crystalloid infusion volume, gastric drainage, and ascites removal. Among seven ML algorithms evaluated, the random forest (RF) model demonstrated the highest predictive performance (AUC 0.731, accuracy 71.6%, sensitivity 62.1%, specificity 76.1%), outperforming conventional scoring systems like sequential organ failure assessment. Importantly, the model uses real-time intraoperative data, making it feasible for intraoperative risk stratification. However, its limitations include a retrospective single-centre design, moderate sensitivity, and the exclusion of preoperative sepsis cases. The authors propose broader validation and have developed an online risk calculator to support clinical deployment[27].

Similarly, Luo et al[28] investigated ML for predicting severe pneumonia after KT, a complication with a 20.9% mortality rate despite prophylactic antimicrobial regimens. Their RF classifier, utilising eight key clinical variables, including preoperative pulmonary lesions, reoperation, age, rabbit anti-thymocyte globulin dose, albumin and immunoglobulin levels, and DGF, achieved a high AUROC of 0.91 and a positive predictive value of 0.85. These features are clinically significant; for example, low immunoglobulin and albumin levels suggest immunocompromise, while DGF and reoperation reflect surgical or graft-related complications. Upon diagnosis, the study outlines a multidisciplinary escalation protocol involving broad-spectrum antimicrobials, immunosuppressive modification, and infectious disease consultation. While the RF model outperformed traditional logistic regression, the study’s retrospective and single-centre design, small sample size, and class imbalance limit external applicability. Future prospective studies are required to validate these promising findings[28].

In paediatric transplantation, Tan et al[29] proposed a predictive model for early graft function recovery in paediatric LDLT. They introduced the dynamic ratio of hepatic function as a novel metric for evaluating graft recovery and incorporated four key risk factors, donor liver insufficiency, prolonged ischaemia time, elevated preoperative creatinine, and high postoperative bilirubin, into a nomogram for clinical use. This tool facilitates early identification of high-risk patients, enabling timely therapeutic interventions such as biopsy or immunosuppressive adjustment. Despite its potential clinical utility, the model's single-centre scope and limited sample size warrant multicentre validation to establish generalisability and long-term prognostic value[29].

ML-based approaches show considerable promise in anticipating serious postoperative complications and graft dysfunction, but their current role is largely exploratory and constrained by methodological limitations. Across liver, kidney, and paediatric transplantation, models such as Random Forest classifiers and nomogram-based tools have consistently outperformed traditional risk scores by integrating preoperative and intraoperative parameters into dynamic predictions. These methods enable earlier recognition of sepsis, pneumonia, and delayed graft recovery, with potential to trigger timely escalation of care or immunosuppressive adjustment. Importantly, they highlight clinically relevant predictors, such as transfusion burden, reoperation, donor liver insufficiency, and bilirubin trajectories, that are mechanistically plausible and actionable. However, the literature remains dominated by retrospective, single-centre designs with modest sample sizes, raising concerns over generalisability and susceptibility to bias. Furthermore, sensitivity and predictive precision, while improved, remain insufficient for independent clinical decision-making. The collective evidence suggests that these models could function as adjunctive tools to augment clinician judgment rather than replacements, with future progress hinging on multicentre validation, balanced datasets, and prospective trials that demonstrate real-world utility and outcome improvement.

Longitudinal monitoring of recovery

Long-term graft function and native renal recovery after transplantation are influenced by complex, evolving variables that necessitate longitudinal assessment. Recent studies have emphasised the value of early post-transplant metrics and preoperative indicators in predicting long-term outcomes, which could guide timely and individualised therapeutic interventions.

Wan et al[30] introduced the recipient-to-donor estimated glomerular filtration rate (eGFR) ratio (RFR) at 3 months post-KT as a prognostic biomarker for long-term graft outcomes. An RFR ≥ 1, indicating recipient renal function equal to or exceeding that of the donor, was associated with an approximately 45% reduction in the 10-year risk of death-censored graft failure compared to an RFR < 1. Optimal graft survival was noted in patients with RFR values between 1 and 2; values outside this range suggested either inadequate recovery (RFR < 1) or possible hyperfiltration injury (RFR > 2), the latter potentially due to donor-recipient size mismatch. Clinical correlates of suboptimal RFR included DGF, early AR, older recipient age, female gender, and use of extended criteria donor kidneys, all reflective of recovery potential after ischaemia-reperfusion injury. While three-month eGFR is a known predictor, RFR offers donor-adjusted insights that can flag patients who may benefit from early interventions like biopsy or immunosuppressive modification. However, the study’s single-centre scope and limited sample size constrain generalisability and call for multicentre validation[30].

Campbell et al[31] provided further insight into native renal recovery following orthotopic liver transplantation alone (OLTa) by emphasising the duration, rather than aetiology, of pretransplant renal dysfunction as a dominant predictor of 12-month post-transplant renal function. Contrary to prior assumptions, the presence of hepatorenal syndrome did not significantly influence recovery. Instead, patients with shorter durations of renal dysfunction, irrespective of cause, had more favourable outcomes. Notably, those undergoing combined kidney-liver transplantation (CKLT) despite worse baseline characteristics (longer dysfunction, higher peak creatinine) experienced better recovery than OLTa recipients with similar profiles, suggesting effective preoperative selection. A duration threshold of 3.6 weeks optimally predicted persistent renal dysfunction (creatinine ≥ 1.5 mg/dL at 12 months), though its predictive value (AUC = 0.71) was moderate. Hence, the authors recommend multifactorial evaluation, including renal replacement therapy (RRT) requirement, MELD score, and imaging findings, to improve CKLT candidacy assessments, acknowledging limitations of creatinine-based GFR estimates in cirrhosis[31].

Levitsky et al[32] advanced this discussion by using radionuclide scans to separate native from transplant kidney GFR in recipients of simultaneous liver-KT (SLK), uncovering wide variability in native renal recovery. This approach highlighted the inadequacy of current criteria, such as the united network for organ sharing (UNOS)-defined native GFR < 20 mL/minute, which, while more predictive than institutional alternatives, still lacked sufficient sensitivity and specificity. Notably, abnormal renal imaging was the only preoperative factor significantly associated with lack of native kidney recovery, underscoring its neglected value in SLK eligibility assessments. Additionally, exploratory plasma biomarkers (e.g., trefoil factor 3, osteopontin, α1-microglobulin, clusterin) showed potential to discriminate reversible from irreversible injury, but their reliability is limited unless serially monitored during the perioperative window. The study also identified poorer outcomes when extended criteria donor kidneys were used for SLK, manifesting as lower GFR, greater RRT dependence, and reduced survival, prompting reconsideration of this practice when native recovery is unlikely and robust renal function is essential[32].

Longitudinal monitoring studies highlight the importance of dynamic and donor-adjusted metrics in predicting graft survival and renal recovery, but also reveal persistent challenges in standardising assessment. The introduction of the recipient-to-donor eGFR ratio (RFR) represents an advance beyond crude measures of post-transplant renal function, offering a donor-calibrated biomarker that captures both adequacy of recovery and risks of hyperfiltration injury. Similarly, the identification of duration, rather than cause, of pretransplant renal dysfunction as the dominant determinant of post-transplant recovery shifts emphasis toward early referral and timely organ allocation decisions. These findings underscore that static measures such as creatinine or conventional UNOS thresholds may be insufficient to capture the nuanced recovery trajectories seen in simultaneous or sequential liver-kidney transplantation. Novel tools such as radionuclide-based separation of native vs graft function and exploratory plasma biomarkers further expand the potential for precision monitoring, though their reliability remains unproven outside small, single-centre cohorts. Taken together, the literature suggests that integrating time-sensitive metrics, imaging, and biomarker data into composite, dynamic models could transform recovery prediction, but robust multicentre validation and consensus on thresholds for intervention are needed before these approaches can meaningfully influence selection criteria or long-term management strategies.

Functionality and rehabilitation

Postoperative functional recovery, particularly of renal function in living kidney donors, is a critical yet underexplored dimension of transplant outcomes. Emerging evidence suggests that perioperative strategies targeting haemodynamic stability and neurohumoral modulation may significantly influence recovery trajectories. Park et al[33] examined the impact of analgesic technique on early remnant kidney function following nephrectomy in living donors, highlighting the role of intrathecal morphine block (ITMB) as a protective perioperative intervention.

Donors who received ITMB demonstrated significantly better early postoperative renal recovery, with a markedly reduced risk, by 74.3%, of delayed functional recovery (defined as eGFR < 60 mL/minute/1.73 m²) on postoperative day 1. In multivariable analysis, male sex, advanced donor age, absence of ITMB, and lower intraoperative fluid administration were all independently associated with delayed remnant kidney recovery. These findings underscore the interplay between analgesic strategy, patient demographics, and intraoperative management in shaping functional outcomes[33].

Mechanistically, ITMB may enhance renal recovery through modulation of the perioperative sympathetic stress response, thereby stabilising systemic and renal haemodynamics. This likely results in improved renal perfusion and reduced ischaemic injury to the remaining kidney. While the hypothesis is biologically plausible and supported by indirect clinical data, direct mechanistic studies confirming the renal-specific haemodynamic effects of ITMB are still lacking[33].

Furthermore, male and older donors consistently exhibited worse recovery profiles. These trends may reflect differences in hormonal milieu, preexisting vascular or renal comorbidities, and reduced nephron reserve associated with aging. Such demographic disparities underscore the need for individualised donor management protocols[33].

Importantly, the study also emphasised the role of intraoperative fluid strategy. Lower fluid infusion rates were associated with increased risk of early postoperative renal dysfunction, highlighting the need to avoid hypovolemia-induced renal hypoperfusion while maintaining caution against fluid overload. Optimising intraoperative volume management remains a critical but often overlooked determinant of renal rehabilitation[33].

Collectively, these observations highlight that postoperative renal functional recovery in living kidney donors is a multifactorial process influenced not only by inherent donor characteristics but also by perioperative management strategies. Analgesic techniques, exemplified by ITMB, appear to exert protective effects through modulation of sympathetic stress and preservation of renal perfusion, suggesting that targeted neurohumoral interventions may meaningfully accelerate early renal rehabilitation. Simultaneously, demographic factors such as sex and age consistently emerge as independent determinants of recovery, likely reflecting baseline nephron reserve, comorbid vascular physiology, and hormonal influences, underscoring the need for individualised perioperative planning. Intraoperative fluid administration further modulates outcomes, with both under- and over-resuscitation carrying risks for remnant kidney dysfunction, indicating that fine-tuned haemodynamic management is essential. Taken together, these insights suggest that functional rehabilitation after donor nephrectomy is best understood through an integrative framework that considers patient-specific risk factors, surgical and anaesthetic strategies, and meticulous perioperative haemodynamic optimisation, rather than through isolated interventions.

Predictive models of outcomes and survival

Advancements in preoperative risk stratification have increasingly focused on leveraging predictive models to guide clinical decision-making, resource allocation, and patient counselling. Two recent studies - by Molinari et al[34] and Sliwinski et al[35] - highlight the evolving role of data-driven tools in forecasting short- and long-term surgical outcomes, particularly in the context of LT and high-risk abdominal or thoracic procedures.

Molinari et al[34] developed a perioperative mortality prediction model for cadaveric LT candidates using readily available preoperative clinical and demographic variables. This model demonstrated exceptional discriminative performance, achieving a C-statistic of up to 0.95 in patients with a predicted ≥ 10% risk of 90-day mortality. In contrast to earlier models that depended heavily on intraoperative or postoperative variables and yielded only moderate accuracy (C-statistics approximately 0.63-0.70), Molinari’s model can be applied at the time of transplant referral, allowing for earlier and more informed decision-making.

Critically, the model deliberately excludes diagnosis and sex to mitigate systemic biases and promote equitable candidate selection. It offers a pragmatic framework for preoperative risk assessment without supplanting clinical judgment. When externally validated, the model could facilitate transparent and standardised assessments of mortality risk, enhancing the selection process, optimising organ allocation, and improving prediction of both short- and long-term outcomes, including 90-day and 1- to 5-year survival. Nonetheless, the study's retrospective design, omission of frailty and nutritional variables, and absence of intraoperative detail represent notable limitations requiring further refinement and validation[34].

In a parallel effort to operationalise preoperative risk assessment, Sliwinski et al[35] piloted the Prehab App, a digital risk calculator designed to identify surgical candidates at elevated risk of poor outcomes using simple, validated metrics, namely the Risk Analysis Index, Eastern Cooperative Oncology Group score, Timed Up and Go test, and haemoglobin level. The tool demonstrated a strong correlation with 90-day survival after major abdominal and thoracic surgeries and required less than five minutes to complete, overcoming many logistical barriers to implementation in routine clinical settings[35].

The App’s utility extends beyond risk stratification; it serves as a gateway for individualised prehabilitation strategies, spanning physical conditioning, nutritional optimisation, and psychosocial support. These interventions are expected to mitigate risk in high-vulnerability cohorts. Although limited by single-centre design and the exclusion of cardiac and orthopaedic cases, the study lays the groundwork for broader implementation. Planned randomised trials and AI-driven enhancements aim to improve predictive precision, enabling modular adaptation across surgical specialties, including hepatectomy, pelvic surgery, and sarcopenia management[35].

Taken together, these studies illustrate a clear trend toward the integration of robust, data-driven predictive models into preoperative risk stratification, with the dual goals of improving patient selection and guiding perioperative management. High-performing models, such as Molinari et al’s perioperative mortality tool[34], demonstrate that carefully curated clinical and demographic variables can achieve exceptional discriminative accuracy, enabling clinicians to anticipate outcomes before surgery rather than retrospectively, and to do so while mitigating systemic biases. Complementary approaches, exemplified by Sliwinski et al’s Prehab App, highlight the practical utility of rapid[35], point-of-care digital risk assessment, which not only predicts vulnerability but also informs tailored prehabilitation interventions to improve functional reserve and postoperative survival. These findings underscore that predictive modelling is most effective when it combines accuracy, accessibility, and actionable guidance, integrating risk quantification with patient-specific optimisation strategies. The overarching insight is that the future of surgical outcome prediction lies in modular, generalisable tools that bridge evidence-based forecasting with individualised perioperative planning, rather than relying solely on traditional scoring systems or intraoperative data.

CROSS-CUTTING METHODOLOGIES AND ADVANCES

Imaging-based AI applications

AI integrated with advanced imaging modalities is reshaping intraoperative and histopathological assessment in transplantation. By overcoming the limitations of conventional techniques, imaging-based AI applications offer precise, rapid, and standardised evaluations that enhance both intraoperative decision-making and pathological interpretation. Notably, recent studies by Felli et al[36] and Marsh et al[37] illustrate the utility of deep learning in hyperspectral imaging (HSI) and digital pathology, respectively, in assessing organ viability and structural pathology.

Felli et al[36] developed an AI-enhanced hyperspectral imaging (AI-HSI) platform to assess liver viability during hepatic artery occlusion, leveraging over 100 spectral bands across the 500-1000 nm range to capture physiological changes invisible to the human eye. This approach enabled operator-independent, rapid evaluation of hepatic oxygenation using tissue-specific indices (StO₂% for superficial and NIR% for deeper microcirculation). Notably, while oxygenation levels initially recovered after 5 hours of reperfusion, subsequent declines indicated microvascular dysfunction, later corroborated by histopathological findings and CD31 immunostaining.

The viability score derived from the AI model strongly correlated with conventional biological markers such as capillary lactates and the Suzuki score, affirming its predictive value for liver injury. Despite being based on a relatively compact convolutional neural network (CNN) trained on a limited dataset, the model achieved exceptional performance (sensitivity 0.993, specificity 0.997), underscoring the potential of HSI-based deep learning even with small sample sizes. In contrast to standard imaging techniques such as ultrasound, CT, indocyanine green fluorescence, and near-infrared spectroscopy, AI-HSI offers non-invasive, real-time functional mapping with minimal operator dependency. However, current limitations include shallow tissue penetration (3-5 mm), no real-time video capability, and limited discrimination of ischaemia type and timing. Future directions include the integration of spectroscopic probes and HYPER-enhanced reality tools to expand clinical applicability, particularly in complex intraoperative scenarios[36].

Complementing this intraoperative innovation, Marsh et al[37] introduced a fully convolutional deep learning model for automated glomerulosclerosis quantification in kidney biopsies, improving both diagnostic speed and consistency. The model incorporated dilated convolution layers to expand the receptive field, allowing for more accurate pixel-wise classification and improved contextual interpretation of glomerular regions. This design enabled superior prediction in areas with large fibrous capsules by reducing false positives near vessels or cystic structures[37].

Crucially, the fully convolutional architecture outperformed traditional patch-based models, offering a 22-fold reduction in processing time for whole-slide images, thus enhancing its feasibility in routine histopathology workflows. Accuracy was further optimised by using overlapping input patches and selecting only central regions for predictions, mitigating edge artifacts. Although some misclassification persisted, particularly around structures mimicking glomeruli, augmenting training datasets with additional tissue types and refining post-processing techniques (e.g., advanced blob detection) could further improve performance. Importantly, this approach reduced inter-observer variability and aligned closely with pathologist assessments, highlighting its potential role in standardising kidney graft evaluation and reducing diagnostic discrepancies[37].

Overall, these studies demonstrate that AI integration with advanced imaging modalities is fundamentally transforming organ assessment in transplantation by enabling rapid, objective, and standardised evaluation of both intraoperative and histopathological parameters. AI-enhanced hyperspectral imaging exemplifies how ML can detect subtle, clinically relevant physiological changes such as microvascular dysfunction or oxygenation deficits that elude conventional imaging, providing real-time, operator-independent insight into organ viability. Deep learning models in digital pathology, such as automated glomerulosclerosis quantification, similarly streamline complex histological analysis, reduce inter-observer variability, and accelerate workflow without sacrificing accuracy. A natural and likely future frontier is the integration of these imaging and histopathology pipelines with molecular biomarkers and multi-omic data (Table 2). Transcriptomic signatures of rejection (for example, the molecular microscope diagnostic system in kidney biopsies), donor-derived cell-free DNA, urinary chemokines such as CXCL10, proteomic and metabolomic profiles of perfusate during machine preservation, and graft tissue spatial transcriptomics each capture biological information that is orthogonal to morphology. Multimodal models that fuse whole-slide images, hyperspectral perfusion data, and these molecular layers offer a plausible route to earlier and more specific detection of rejection, ischaemia-reperfusion injury, and recurrent disease than any single modality. Such integration will require harmonised data standards, prospective biobanking with linked clinical outcomes, and methods that handle missingness and high dimensionality without sacrificing interpretability. In summary, imaging-based AI is shifting the paradigm from subjective, labour-intensive interpretation toward reproducible, data-driven decision-making, with the dual potential to optimise graft selection, guide intraoperative management, and standardise post-transplant pathological evaluation; coupling these advances with molecular and multi-omic readouts is likely to expand clinical applicability and improve predictive precision across transplant settings.

Table 2 Consolidated future directions for artificial intelligence and machine learning research across the transplantation surgery care pathway, organised by domain.

Domain	Future directions
Preoperative imaging & planning	Multicentre validation, inclusion of diverse donor/recipient populations, integration of AI tools into clinical workflows, expansion to paediatric and left-lobe graft planning
Risk stratification & functional status	Prospective multicentre validation, integration of objective frailty and performance measures, incorporation of socio-clinical context into models
Prehabilitation & surgical eligibility	Broader external validation, development of precision prehabilitation programmes, incorporation of metabolic biomarkers, and standardised eligibility frameworks
Perioperative risk prediction	Incorporation of intraoperative real-time signals, multicentre prospective trials, explainability-focused models for clinical adoption
Robotic & ML-enhanced surgery	Cost-effectiveness studies, technological refinement, definition of training standards, long-term comparative outcome studies
Postoperative monitoring	Large-scale validation, integration with real-time EHR systems, multimodal models combining clinical, imaging, and biomarker data
Longitudinal recovery & rehabilitation	Validation of composite recovery biomarkers, mechanistic studies on perioperative interventions, personalised rehabilitation protocols
Cross-cutting imaging & pathology	Development of real-time video-capable imaging, dataset expansion, integration of AI histopathology into diagnostic standards
Multiorgan & multidisciplinary models	Multimodal, cross-organ models; integration into perioperative and critical care decision-support systems; dataset standardisation across institutions
Ethical & regulatory frameworks	Development of ethical governance frameworks, bias audits, and explainable AI standards to guide deployment in transplantation

EHR: Electronic health record; ML: Machine learning; AI: Artificial intelligence.

Multiorgan and multidisciplinary applications

AI is increasingly being applied across organ systems and clinical domains, enabling integrated, data-driven insights that transcend traditional subspecialty boundaries. In the field of nephrology and transplantation, the work of Rashidi and Bihorac exemplifies the breadth of AI applications, spanning predictive analytics, diagnostic imaging, intraoperative risk assessment, and histopathological interpretation[38].

One of the most impactful contributions described by Rashidi and Bihorac is the deployment of deep recurrent neural networks (RNNs) to predict inpatient episodes of AKI using longitudinal EHR data. These models can anticipate AKI onset up to 48 hours in advance, offering a critical window for early intervention in high-risk hospitalised patients. Unlike static risk scores, RNNs dynamically incorporate evolving clinical parameters, allowing real-time risk recalibration as patient physiology changes[38].

Building upon this framework, they further integrated intraoperative physiological signals into AKI risk prediction models. By combining preoperative data with intraoperative time-series metrics (e.g., blood pressure, oxygenation, fluid status), the model provides a continuous, adaptive risk estimate for postoperative AKI. This multimodal approach aligns closely with the perioperative care of transplant and critically ill surgical patients, where real-time decision support can significantly impact outcomes[38].

In diagnostic imaging, CNNs have enabled non-invasive classification of chronic kidney disease (CKD) stages and eGFR using ultrasound images of the kidneys. This represents a shift toward image-based phenotyping of renal function, potentially reducing reliance on serum creatinine and providing a radiation-free alternative to CT or nuclear medicine-based assessments. The model’s ability to discern subtle anatomical changes associated with CKD progression holds particular promise in resource-limited settings or for longitudinal outpatient monitoring[38].

Moreover, CNNs have also been trained for multiclass segmentation of renal histopathology, allowing automated analysis of transplant biopsy and nephrectomy specimens. By segmenting glomeruli, tubules, interstitial spaces, and vasculature, these models facilitate objective quantification of pathologic features, reducing inter-observer variability and accelerating diagnostic workflows. In transplantation, such tools may support standardised grading of rejection, fibrosis, and chronic allograft injury, augmenting both clinical decision-making and research reproducibility[38].

Taken together, these studies illustrate that AI is enabling a truly integrative, cross-domain approach to patient care, bridging preoperative, intraoperative, and post-transplant management while encompassing imaging, histopathology, and longitudinal clinical data. RNNs applied to dynamic EHR datasets exemplify the predictive power of AI for time-sensitive complications, such as AKI, providing clinicians with actionable lead time for early intervention. When combined with intraoperative physiological monitoring, these models offer continuously updated risk estimates, reflecting the complex, evolving perioperative physiology of transplant patients. Concurrently, CNNs in both diagnostic imaging and histopathology facilitate precise, reproducible phenotyping of renal structure and function, reducing reliance on invasive or subjective assessments and standardising evaluation across clinical and research contexts. AI’s strength lies not merely in isolated applications, but in its ability to integrate heterogeneous data streams across organ systems and specialties, supporting proactive, personalised, and evidence-based decision-making throughout the continuum of care.

Future directions and prioritised roadmap

Rather than an exhaustive catalogue, we propose a prioritised roadmap for clinical translation. Priorities are ordered by their likely near-term impact on patient outcomes and on the credibility of AI in transplantation; each requires multicentre collaboration and harmonised data infrastructure.

Prospective multicentre and external validation. The most pressing need is to test existing single-centre models on independent, geographically diverse cohorts, including temporal validation against future patients, before any deployment. Pre-registered protocols and standardised reporting (TRIPOD-AI, DECIDE-AI) should be the norm[39,40].

Dataset diversity, equity, and bias auditing. Models should be trained and tested on cohorts that reflect the populations served, including under-represented ethnic groups, women, older recipients, paediatric patients, and recipients of marginal or DCD grafts. Routine subgroup performance reporting and explicit bias audits should accompany every deployment.

Paediatric and complex-graft datasets. Federated learning across paediatric and split- or left-lobe transplant centres is required to overcome small sample sizes and to develop graft-specific, age-adjusted models that adult-derived algorithms cannot reliably provide.

Integration of continuous intraoperative and longitudinal physiological signals. Coupling ML with high-frequency anaesthetic record data, machine perfusion parameters, and wearable-derived activity profiles will enable dynamic rather than snapshot risk estimation throughout the pathway.

Multi-omic and multimodal fusion. Combining imaging, histopathology, donor-derived cell-free DNA, urinary biomarkers, and transcriptomic signatures (for example, the molecular microscope diagnostic system) in unified models is a credible route to earlier and more specific detection of rejection and ischaemia-reperfusion injury.

Explainable AI and clinical decision support. Real-time decision support should expose feature attributions (for example, SHAP values) directly within the electronic anaesthetic and EHR interface, with clear alerting thresholds and prospective evaluation of clinician response, not merely predictive performance.

Health technology assessment and ethical governance. Cost-effectiveness analyses of robotic and AI-assisted platforms, alongside formal regulatory pathways for adaptive ML systems, are required to support equitable adoption and reimbursement; ethical oversight, including auditable model registries and post-deployment surveillance, should be embedded from the outset.

CONCLUSION

AI and ML are moving from proof-of-concept toward genuine decision support across the transplantation pathway, with the strongest evidence in preoperative imaging and volumetry, intraoperative risk prediction, and digital pathology. Most published models, however, remain retrospective, single-centre, and trained on narrow donor and recipient demographics, so claims of generalisability outpace the evidence. Their responsible deployment requires rigorous prospective validation, dataset diversity (with explicit attention to paediatric and complex grafts), explainable outputs, bias auditing, and integration into clinical workflows rather than parallel screens. The agenda set out in the roadmap above provides a pragmatic starting point for clinicians, researchers, and regulators seeking to translate these tools into reproducible improvements in graft survival, patient outcomes, and organ utilisation.