Bladder cancer is not usually discovered early. In many health systems, the route to diagnosis begins when a patient notices blood in their urine and seeks medical advice. That symptom is alarming but far from specific: kidney stones, infections and prostate conditions can produce the same sign. Investigating it typically means a cystoscopy, a procedure in which a thin camera is passed into the bladder. The result is a diagnostic pathway that is both expensive for health services and unpleasant for patients, with many investigations finding nothing serious. Researchers at the University of Plymouth have built an artificial intelligence system that tries to work earlier and more selectively. It reads the long history contained in electronic health records and looks for patterns that precede a bladder cancer diagnosis. Reported in IEEE Transactions on Biomedical Engineering, the system identified most eventual cases and correctly cleared most people who never developed the disease. The team stresses that this is a research result, not yet a clinical service.

Why blood in urine is an imperfect signal

Current referral guidance for suspected bladder cancer leans heavily on visible blood in urine, known medically as haematuria. It is a sensible rule in one respect: blood is a warning sign that deserves investigation. The difficulty is that the same warning can have many benign explanations.

Kidney stones can irritate the urinary tract and cause bleeding. Prostate problems in men can do the same. Infections, injuries and other conditions may produce a similar picture. A patient who presents with blood in urine is therefore entering a broad differential diagnosis, and only a fraction will turn out to have cancer.

The confirmatory procedure adds another layer of difficulty. Cystoscopy allows a clinician to inspect the bladder lining directly, but it is uncomfortable and carries small risks. Health services cannot justify performing it on everyone with a vague urinary symptom, so referral thresholds are set carefully.

That balancing act creates the central problem. If the threshold is too loose, clinics fill with investigations that find nothing. If it is too tight, some cancers are missed until they are more advanced. Anything that can sharpen the decision before referral would be useful to both patients and clinicians.

The Plymouth study and its dataset

The study was led by the University of Plymouth and published in IEEE Transactions on Biomedical Engineering, with the DOI 10.1109/TBME.2026.3658230. The paper's title describes early detection of bladder cancer using advanced feature engineering and swarm intelligence optimisation on electronic health records. Its author list includes Xu Wang, Andrew Preston, Josiah Aning, Miklos Loizou and Shang-Ming Zhou.

The data came from the SAIL Databank in Wales, a secure environment that links anonymised health records for research. The team drew on records spanning 1995 to 2020, covering nearly 70,000 patients. That length of history is important, because the signals the model seeks may appear long before a formal diagnosis.

The model itself is named PRECISE-AGZ. It belongs to a family of approaches that treat a patient's record as a sequence of events: diagnoses, prescriptions, test results, referrals and demographic facts. Machine learning systems can look for combinations of those events that are more common in people who later receive a cancer diagnosis.

The University of Plymouth released a public account of the work on 25 August 2026. ScienceDaily reported it on 21 November 2026. Both accounts emphasise the same caveat: the system shows promise but needs validation in other populations before it could guide referrals.

Mining decades of Welsh health records

Electronic health records were not designed with machine learning in mind. They are administrative and clinical tools, assembled over years from many sources. A single patient's record may contain coded diagnoses, free-text notes, medication histories and laboratory results, with gaps and inconsistencies along the way.

For a researcher, that messiness is both a problem and an opportunity. The records capture what actually happened to patients in routine care, not just what happened in a controlled trial. They can reveal patterns that a clinician, seeing one patient at a time, would find hard to notice.

The Welsh dataset offered a particular advantage. SAIL links records across settings, which means a patient's interactions with different parts of the health system can be assembled into a coherent timeline. The team could look backwards from a diagnosis date and ask what had happened in the years before.

Using data from 1995 to 2020 also means the study spans changes in coding practices and referral guidelines. Any system trained on such data must be robust to those shifts, and the authors acknowledge that validation in a different time period and a different health system is essential.

From 48,261 indicators to 38 features

The scale of the search is one of the study's most striking details. The team began with 48,261 candidate indicators drawn from the health records. These included diagnoses, medications, clinical activities and other events that might conceivably carry information about bladder cancer risk.

A model given tens of thousands of inputs would be difficult to train and even harder to interpret. Many of those indicators would be irrelevant, and many others would be redundant, appearing in almost identical patterns. The researchers therefore used feature engineering and optimisation methods to narrow the field to 38 features.

That reduction is not simply a convenience. A smaller set of features makes the model more transparent, cheaper to run and less likely to latch onto artefacts of the dataset. It also makes the result easier to explain to clinicians, who need to understand why a system is flagging a patient.

The study's use of swarm intelligence optimisation refers to search methods inspired by collective behaviour, such as how flocks or colonies find solutions. In this context, the method helps identify a useful combination of features from a very large set of possibilities. The paper's title signals that this was a central part of the approach.

A compact feature set has another advantage. Clinicians are more likely to trust a system whose reasoning can be described in familiar terms, such as recent diagnoses, prescriptions and investigations, than one that draws on thousands of opaque indicators. The researchers present the 38 features as part of the model's design, although they do not claim that each one is a cause of cancer.

The headline results: 85 and 91 percent

The reported performance figures are straightforward to state. The model identified 85 percent of people who had bladder cancer, a measure known as sensitivity. It also correctly classified 91 percent of people who did not have the disease, known as specificity.

Those numbers matter together. A system with very high sensitivity that flagged almost everyone would be useless in practice, because clinics would still be overwhelmed. A system with very high specificity that missed most cancers would be unsafe. The reported combination suggests a useful balance, at least within the dataset used.

ScienceDaily reported that the AI performed better than existing NHS referral guidance in this analysis. That comparison is meaningful because it relates to the real decision clinicians face: who should be referred for further investigation. The study does not claim that the system has been tested prospectively in clinics.

The authors also make clear that these figures come from a research dataset. Performance can fall when a model meets new hospitals, new coding practices and new populations. That is why the team describes wider validation as a necessary next step rather than a formality.

Three risk bands for triage

Rather than producing only a single yes-or-no answer, the system sorts patients into three risk bands. The University of Plymouth release describes bands below 7 percent, between 7 and 55 percent, and above 55 percent probability of bladder cancer.

That design reflects how triage works in practice. A patient in the lowest band might reasonably be managed without urgent referral. A patient in the highest band would be a strong candidate for prompt investigation. The middle band is where clinical judgement and further testing would remain most important.

Graded output is also more honest than a binary classification. Most medical predictions are uncertain, and presenting that uncertainty helps clinicians use the result appropriately. A probability estimate allows a doctor to weigh the model's output against symptoms, examination findings and patient preferences.

The bands are a proposal, not an established clinical standard. Before they could be used in a health service, researchers would need to show that patients placed in each band really do have the stated risk in a new population, and that acting on the bands improves outcomes.

Signals that appear years before diagnosis

The most attention-grabbing finding is timing. The model found warning signs a year before diagnosis in many cases, and in some patients the signals stretched back as far as five years. That is much earlier than the current symptom-led pathway typically allows.

Early detection matters because bladder cancer outcomes depend heavily on the stage at diagnosis. A tumour found while it is still superficial is generally more treatable than one that has invaded deeper tissue or spread. Any tool that moves diagnosis earlier could, in principle, improve survival and reduce treatment burden.

The pattern also illustrates why electronic records are valuable. A single visit may show nothing remarkable. A sequence of small events over several years can be informative when viewed as a whole, even if no individual event would trigger concern.

It is important to be precise about what "warning signs" means here. These are statistical associations in historical records, not symptoms that a patient would necessarily notice. The model is detecting changes in clinical activity and coded information, some of which may reflect subtle changes in a patient's condition before diagnosis.

Surprising associations and their limits

The system surfaced associations that were not part of conventional bladder cancer thinking. The Plymouth release mentions links involving dementia, Parkinson's disease and the use of tamoxifen, a drug used in breast cancer treatment.

Findings like these are exactly what data-driven methods are good at producing. They can identify correlations that a hypothesis-driven study might never think to test. They are also easy to misinterpret. An association in a health record does not show that one condition causes another, or that a drug increases cancer risk.

There are several mundane explanations. Patients with dementia or Parkinson's disease may interact with the health system in ways that generate different record patterns. Tamoxifen use may be a marker for a patient group with different screening histories. The authors treat these associations as questions for further study, not conclusions.

That caution is a strength of the paper. Medical AI has a poor record when systems are deployed on the basis of correlations that nobody has examined carefully. Flagging unusual features for investigation, while refusing to overclaim, is the appropriate response.

What validation would require

The next step is testing in other datasets and health systems. A model trained on Welsh records from 1995 to 2020 may depend on features of that particular setting: coding conventions, referral patterns, prescribing habits and the structure of the SAIL Databank itself. Validation would show which parts of the model travel.

Researchers would also need to test performance in a prospective study, where the model's predictions are generated in real time and compared with what actually happens. Retrospective results can be optimistic because the data used to build the model are also the data used to score it, unless careful separation is enforced.

Clinical validation goes beyond accuracy. A health service would need to know whether using the system changes practice in a useful way. Does it increase the proportion of cancers found at an early stage? Does it reduce unnecessary cystoscopies? Does it delay diagnosis for anyone who would have been referred under existing guidance?

Equity is another concern. A model trained on one population may perform differently for older patients, women, ethnic minorities or people with comorbidities. Checking performance across groups would be essential before any deployment.

Calibration would matter too. A model can rank patients correctly yet still quote probabilities that are too high or too low, which would distort referral decisions. Researchers would need to check whether a quoted risk of 40 percent really corresponds to a 40 percent chance in a new population. They would also need to show that acting on the predictions produces more benefit than harm, a question that accuracy figures alone cannot answer.

From research tool to clinical practice

If those validation steps succeed, the system could eventually sit within existing referral pathways rather than replacing them. A general practitioner might use it to refine a decision about a patient with ambiguous urinary symptoms, particularly where the result would change whether a referral is made.

Integration is not trivial. Electronic health record systems differ between hospitals and countries, and a model that depends on 38 specific features needs those features to be available reliably. Software that works in a research environment may not transfer cleanly into a busy clinic.

There is also a human factor. Clinicians need to understand how much to trust a probability estimate and what to do with it. Training and clear explanation would be as important as the algorithm itself. A system that produces numbers nobody knows how to interpret will not change care.

The paper is therefore an early stage of a longer process. It establishes that the signal exists in routine data and that a model can extract it. Turning that into a service requires engineering, evaluation and regulatory work that this study does not attempt.

What the study does not settle

It would be wrong to read this work as evidence that AI is ready to replace clinical judgement in cancer detection. The model's reported accuracy is impressive, but it comes from a specific dataset analysed after the fact. Many promising medical AI systems have looked strong in such studies and then disappointed in practice.

Nor does the study change the fundamental advice that blood in urine should be investigated. The system is designed to help decide who needs investigation and perhaps to identify people who have not yet developed overt symptoms. It is not a replacement for established pathways.

The broader significance lies in the method. The team showed that a very large set of routine clinical indicators can be reduced to a manageable model that detects a real signal years ahead of diagnosis. That is a proof that the data contain useful information, even if the precise clinical use is still to be determined.

For patients, the hope is earlier diagnosis with fewer invasive investigations. For health services, the appeal is better use of scarce diagnostic capacity. Both remain conditional on validation. The study is a carefully bounded step toward those goals, and its authors are right to present it that way.

Bladder cancer research has often focused on new tests: urine biomarkers, imaging techniques and improved cystoscopy. This study takes a different route by asking what information is already sitting in health records. The answer, according to the Plymouth team, is more than expected.

The numbers are likely to be quoted: nearly 70,000 patients, 48,261 indicators reduced to 38 features, 85 percent sensitivity, 91 percent specificity and signals up to five years before diagnosis. Each describes a real aspect of the work, but none alone captures its cautious character.

The authors repeatedly point to the need for validation, treat surprising associations as hypotheses and stop short of claiming clinical readiness. In a field where AI announcements often outrun evidence, that restraint is worth noting.

If later studies confirm the findings, PRECISE-AGZ could become part of a more anticipatory approach to bladder cancer, one in which routine data prompt investigation before symptoms become obvious. For now, it is a demonstration that the possibility is worth taking seriously.


Corrections and updates

Nexuswild welcomes factual corrections. Email contact@nexuswild.com with evidence and the article URL.