The core problem the researchers set out to solve is a scarcity mismatch. In primary care, general practitioners record diagnostic codes for chronic kidney disease stages, and these GP-recorded labels carry the weight of clinical judgement: they reflect not only laboratory measurements but also the broader context of a patient’s history, comorbidities, and the physician’s overall assessment. The trouble is that such expert annotations are scarce, because coding is inconsistent, incomplete, and dependent on individual practice. On the other side of the ledger sits a far more abundant label source: estimated glomerular filtration rate, or eGFR, the workhorse blood-test-derived measure of kidney function. Simple rule-based systems can automatically assign a disease stage to virtually any patient who has an eGFR reading, producing labels at population scale. But these rule-based labels are mechanical. They capture a single biochemical threshold and may miss the wider clinical picture that a GP weighs when assigning a stage, which means the two label sources are complementary but only partially consistent with one another.

Previous approaches to this dilemma have tended to treat it in binary terms. Either the label sources are assumed to agree, in which case they can simply be pooled into one training set, or they are assumed to conflict, in which case one is trusted and the other discarded or down-weighted. The Swansea team, led by doctoral researcher Ali Guran together with Avishek Siris, Gary K. L. Tam and Xianghua Xie from the Department of Computer Science, with clinical interpretation provided by James Chess of the Renal Unit at Morriston Hospital, took a different view. They argued that agreement between a GP annotation and a rule-based eGFR label is not a yes-or-no question but a matter of degree. Two labels might match exactly, or they might be adjacent stages, or they might diverge by several stages. That graded structure, they reasoned, contains real information about how confident and clinically meaningful any given case is, and a learning algorithm should be able to exploit it.

The vehicle for that idea is contrastive learning, a technique that has transformed fields from computer vision to medical imaging. In a standard contrastive setup, a neural network is trained to pull representations of similar examples close together in an abstract embedding space while pushing dissimilar examples apart. The Swansea framework generalises this binary pull-and-push logic into a hierarchy. Cases where the GP label and the eGFR rule-based label agree perfectly are treated as strongly concordant and drawn into tight clusters. Cases where the two sources disagree by one stage are pushed apart, but less forcefully. Cases with larger discrepancies are separated progressively more. The result is an embedding space with a graded, clinically interpretable geometry: distance in that space corresponds to the degree of discordance between supervision sources, rather than to an arbitrary similarity metric learned from unlabeled statistics alone. This hierarchical contrastive objective runs alongside a conventional classification loss, so the model still learns to predict stages explicitly while simultaneously absorbing the relational structure of its heterogeneous supervision.

Technically, the framework had to be architecture-agnostic, because longitudinal clinical data, sequences of measurements and events recorded over years of patient care, can be modelled in many ways. The researchers evaluated their approach across a broad portfolio of neural backbones: recurrent networks in the LSTM tradition, standard convolutional architectures, temporal convolutional networks that process long sequences with dilated causal filters, Transformer models built on self-attention, and hybrid designs that combine convolutional feature extraction with attention-based sequence modelling. The hierarchical contrastive loss was layered onto each of these in turn, allowing the team to isolate the contribution of the supervision strategy from the choice of architecture. Across the board, the structured contrastive supervision delivered consistent improvements over two natural baselines: classification-only training, which ignores the relational structure entirely, and binary contrastive training, which treats all disagreements as equivalent.

The strongest overall performance came from a hybrid backbone that couples a temporal convolutional network with a Transformer, an architecture that appears well suited to the rhythm of primary-care records, where long stretches of routine measurements are punctuated by clinically dense episodes. The consistency of the gains across architectures is arguably the most important finding, because it suggests the benefit flows from the supervision design itself rather than from a lucky interaction with one particular model family. For practitioners, that means the hierarchical contrastive recipe is a portable tool: it can be dropped onto whatever sequence model a research group already uses, and it asks only for the two label sources that health systems already generate as a by-product of care.

The study was conducted on data from the Secure Anonymised Information Linkage Databank, a national privacy-protecting research resource in Wales that links primary-care, hospital and other health records for millions of people. All analysis took place inside the SAIL Trusted Research Environment, with the project approved by the databank’s independent Information Governance Review Panel under project number 1220, and only disclosure-controlled aggregate outputs leaving the secure environment. Because the researchers worked exclusively with anonymised, routinely collected data and never accessed identifiable patient information, individual informed consent was not required under the applicable governance framework. This kind of privacy-preserving infrastructure is increasingly the template for large-scale clinical machine learning, and the study demonstrates that such environments can support sophisticated deep learning workflows without compromising patient confidentiality.

The implications reach well beyond nephrology. The mismatch the team addressed, a small reservoir of rich expert labels sitting alongside an ocean of cheap, rule-derived labels, is endemic in medicine. Oncology registries, cardiovascular coding, respiratory diagnostics: nearly every specialty has its own version of the same asymmetry. By showing that graded agreement between label sources can be converted into a training signal, the Swansea framework offers a general template for longitudinal disease modelling under heterogeneous supervision. The authors note that the approach could transfer to other chronic conditions where clinician judgement and algorithmic rules coexist, and where the gap between them is itself informative. In effect, the method treats the healthcare system’s own inconsistencies as a feature to be learned from, rather than a defect to be cleaned away.

There are, of course, caveats. The underlying data cannot be shared publicly, and code components tied to the secure environment are available only on reasonable request subject to governance review, which means independent replication will require comparable trusted-research arrangements. The rule-based eGFR labels themselves inherit the limitations of the KDIGO-style staging criteria from which they derive, and the framework does not resolve which source is ‘right’ when they disagree; instead it encodes the disagreement as structure. But that reframing is precisely the contribution. Rather than forcing a choice between clinical nuance and population scale, the study shows how to have both, and it does so with an elegance that machine learning researchers will recognise immediately: when your labels disagree, teach your model the shape of the disagreement. For a disease that silently erodes kidney function in roughly one in ten adults worldwide, tools that squeeze more signal from the records health systems already collect could translate into earlier detection, better staging, and ultimately better outcomes for patients who today slip through the diagnostic cracks.

Subject of Research: Hierarchical contrastive learning combining GP-recorded and eGFR labels for chronic kidney disease staging

Article Title: Hierarchical contrastive learning from GP-recorded and eGFR labels for CKD staging

Article References: Guran, A., Siris, A., Tam, G. K. L., Chess, J., & Xie, X. (2026). Hierarchical contrastive learning from GP-recorded and eGFR labels for CKD staging. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03668-z

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03668-z

Keywords: chronic kidney disease, contrastive learning, machine learning, eGFR, clinical labels, longitudinal data, Transformer, temporal convolutional network, SAIL Databank, primary care, deep learning, disease staging