Why There Is No Single AI for Transplantation: The Two Questions That Decode Every Algorithm

Consider an ordinary moment in a transplant clinic. A kidney transplant recipient arrives for routine follow-up with an elevated serum creatinine. Before reaching any conclusion about possible rejection or injury, the clinician does something remarkable without noticing: they look at the trend of the value rather than a single reading, check the proteinuria, recall the donor-specific antibodies measured last year, open the most recent biopsy report, and weigh all of it against everything known about this particular patient. In a few minutes, five different kinds of information have been fused into one judgement. That ordinary act of synthesis, researchers now argue, turns out to be the single hardest thing for artificial intelligence to replicate, and understanding why reframes the entire debate about AI in transplantation.

A new perspective published in Transplant International, launching an educational series called Decoding AI for Transplantation, challenges the way clinicians, journalists and patients talk about medical AI. The authors, a team spanning clinical transplantation, biostatistics and surgery in Paris, Madrid, Leuven and Vienna, argue that the pervasive habit of evaluating artificial intelligence as a single monolithic entity, a uniform black box expected to revolutionize care overnight, is fundamentally misleading. In reality, there is no singular all-encompassing AI tool for transplantation. Instead, the field consists of many highly narrow algorithms, each engineered for a specific purpose and each reading a particular kind of data. The essential clinical skill, they contend, is not the capacity to critique AI in the abstract but the practical ability to tell these disparate technologies apart.

The framework the authors propose rests on two deceptively simple questions. The first concerns modality: what kind of data does the tool actually read? Some tools consume structured clinical data such as creatinine values, proteinuria measurements, HLA mismatch counts and donor-specific antibodies. Others read images, typically digitized biopsy slides. Some read molecular data, whether the gene-expression signature of a tissue sample or donor-derived cell-free DNA circulating in the blood. Still others follow longitudinal data, tracking how a value moves over months rather than its level on a single day, and a growing class of tools now reads free text, including clinical notes, referral letters and the scientific literature, using large language models. Each modality is, in effect, a different language, and most tools speak only one of them.

The second question concerns the task: what clinical job is the tool built to perform? A tool may be designed to diagnose what is happening in the graft right now, to screen or monitor for something developing before it becomes clinically obvious, to predict long-term risk, or to guide a treatment decision. The same modality can serve different jobs, and the same job can be approached through entirely different modalities, which is precisely why the word AI on its own conveys so little. Placing real transplant tools on these two axes makes the point concrete. Digital-pathology models read biopsy images to help diagnose rejection. Transcriptomic classifiers, such as those built on the Banff Human Organ Transplant gene panel and the MMDx molecular phenotyping project, read molecular data to answer the same diagnostic question by a completely different route. Donor-derived cell-free DNA assays read the blood to monitor for subclinical rejection before a biopsy is ever taken.

One tool in the landscape already stands out for combining sources rather than relying on a single stream. The iBox risk score, developed through an international derivation and validation study, reads clinical, histological and immunological data together to predict long-term graft survival after kidney transplantation. Four tools, four distinct positions on the modality-by-task map, and only by naming each tool’s data input and clinical objective can a clinician judge what it genuinely offers and, just as importantly, what it cannot. This categorization, the authors argue, should become the reflexive first move whenever a new AI product or paper arrives in the transplant community.

Categorizing a tool this way exposes a critical limitation that a single headline metric can easily hide: most algorithms capture only a narrow slice of the total clinical picture. The biopsy-reading model has never seen the cell-free DNA result. The purely clinical risk score knows nothing of the histology. Each model can be excellent within its specific diagnostic domain and is, by design, blind to everything outside it. That is the exact reverse of what a clinician does at the bedside, where the value of the judgement comes not from any one number but from holding individual results together, the rising creatinine made meaningful by the proteinuria, the antibodies reinterpreted in the light of the biopsy findings.

A narrow answer delivered with confidence, the authors warn, is easy to mistake for a complete one. A tool that returns one confident result from one source is doing something a clinician cannot, answering a narrower question often very well, but a narrower one nonetheless. To be fair, some tools already reach beyond a single modality, and the honest description of the field is a spectrum: many single-modality tools, a growing few that combine two or three sources, and none that yet sees the whole patient. The authors are careful to note that neither does the clinician alone, and that multimodal integration is precisely the path toward seeing more than either the human eye or any single tool can on its own.

The frontier of the field, and the reason many promising tools stumble, lies in making that integration work. Different modalities live on different scales and timeframes: a histology slide, a single blood value and a years-long trajectory of lab results are not naturally comparable, and forcing them into one model confronts researchers with a genuine technical problem rather than a formality. Sources are often missing in clinical practice, with no biopsy performed at a given visit or an anti-HLA donor-specific antibody result not yet back from the laboratory, and a model trained on complete data can behave unpredictably when inputs are absent. Unlike traditional statistics, which offers a mature and flexible arsenal of methods for handling missing data, AI methodology has not yet reached that stage. Moreover, the more sources a model blends, the harder it becomes to see why it reached its conclusion, exactly at the moment when clinicians most want to know.

This is where foundation models and large language models enter the picture, and the authors are careful about their role. These general-purpose systems, often trained on vast text corpora, are increasingly proposed not as oracles that replace clinical judgement but as connective tissue, able to take in very different kinds of input and help bind them into a coherent whole. It is a useful job and a difficult one. The stated ambition, which the authors call the holy grail, would be to train computational models to replicate what clinicians already perform instinctively: evaluating the patient holistically rather than viewing an isolated organ system or a singular assay. Yet the authors argue the goal should not be to develop machine learning methods that replicate what clinicians do, but rather to provide better tools that help clinicians make better decisions.

For clinicians and researchers confronting the growing flood of AI claims, the authors distill the reframing into five questions that turn an intimidating field into a set of answerable ones. What kind of data does the tool read? What is its clinical job? Does it read one modality or combine several, and if it integrates, how does it cope when a source is missing? Was it tested in patients and centres other than the ones it was built on, the crucial test of external validation? And does it actually change a decision or improve an outcome, or does it only describe a risk the clinician already sensed? A tool that reads one modality for one task can be excellent and still be only part of the picture. A tool that claims to integrate many should be held to a higher standard, not a lower one, because its reach is wider and its failures quieter. The question, the authors conclude, is not whether AI can replace clinical judgement but how it can be used to make the best possible judgement to help patients.

Subject of Research: A framework for classifying and critically appraising artificial intelligence tools in organ transplantation by data modality and clinical task

Article Title: Decoding AI for transplantation: many models, one patient: why there is no single “AI for transplantation” – and what to ask instead

Article References: Aubert, O., Cortes Garcia, E., Pineda, S., Neyens, T., & Pilat, N. (2026). Decoding AI for transplantation: many models, one patient: why there is no single “AI for transplantation” – and what to ask instead. Transplant International, 39, Article 17554. https://doi.org/10.3389/ti.2026.17554

Image Credits: AI Generated

DOI: 10.3389/ti.2026.17554

Keywords: artificial intelligence, transplantation, multimodal AI, kidney allograft, rejection diagnosis, cell-free DNA, digital pathology, foundation models, external validation, clinical decision support, machine learning, iBox risk score