Artificial intelligence-driven kidney organ allocation: systematic review of clinical outcome prediction, ethical frameworks, and decision-making algorithms.
Firuzpour, F., et al.BMC Nephrol. 2025 Nov 14;26(1):639.
Aims
This study aimed to synthesise the available evidence on AI-based kidney allocation, focusing particularly on predictive modeling, allocation algorithms, and ethical considerations.
Interventions
Five electronic databases were searched for relevant literature: PubMed, MEDLINE, Embase, Web of Science and Scopus. AI/computer science repositories and clinical trial registries were also searched. The methological quality of the included studies were searched using modified QUADAS-2 and PROBAST frameworks.
Participants
16 studies were included in the review.
Outcomes
Primary outcomes included graft survival, patient survival, longevity and waitlist mortality. The secondary outcomes were explainability of models, fairness/equity metrics and policy simulation results.
Follow-up
N/A
CET Conclusions
This systematic review tackles an emerging and complex field at the intersection of transplantation, data science, and ethics. The authors conducted an expansive search across biomedical databases, computer science repositories, and trial registries, capturing both clinical and technical literature. PRISMA guidance was followed and a clearly defined PICOS framework was in place. The quality of included studies was assessed using modified QUADAS-2 and PROBAST frameworks. A narrative synthesis is undertaken, which highlights that most models remain retrospective and predictive, with only a minority integrated into simulated or operational allocation frameworks. Commonly employed methods include both classical survival analysis (such as Cox proportional hazards models and decision trees), ensemble techniques (e.g., random survival forests, Cox ensemble), and various deep learning strategies (e.g., DeepSurv, DeepHit, Deep Cox mixture models, neural networks). These models utilize features from both donors and recipients—such as age, comorbidity, HLA mismatch, donor eGFR, cold ischemia time, and transplant centre characteristics—to improve risk stratification and guide decision support. Many reported performance gains over traditional risk scoring systems (C-indices ~0.65–0.72), but this may reflect the number and type of variables used, rather than algorithmic superiority, a point the authors acknowledge. There were some included papers reporting on the integration of AI/ML into allocation systems, and this was evaluated primarily in simulated environments. However, there are no reports of real-world clinical deployment. The limitations of the study primarily reflect the underlying evidence base. Included studies were highly heterogeneous in design, data sources, validation strategies, and performance metrics. External validation was uncommon, prospective evaluation absent, and fairness metrics inconsistently defined and rarely embedded. The authors’ conclusion—that AI shows promise but is not ready for clinical deployment in kidney allocation—is well justified. Their emphasis on governance, transparency, and multidisciplinary validation is to be applauded. Overall, this is a careful and appropriately restrained report, acknowledging the potential in AI/ML without hyperbole.
Trial registration
N/A