Years before bladder cancer is diagnosed, faint warning signs may already be accumulating across a patient’s medical history.
Bladder cancer causes about 220,000 deaths worldwide each year, but finding it early remains difficult. Blood in the urine is its best-known warning sign, yet the same symptom can result from kidney stones, prostate problems, and other noncancerous conditions.
Researchers led by the University of Plymouth have developed an artificial intelligence system that searches electronic health records for combinations of symptoms, risk factors, prescriptions, and clinical activity associated with the disease. Some of these signals appeared as early as five years before diagnosis, although the model performed most effectively within the final 12 months.
Finding Cancer Clues in Health Records
The research team, led by Professor Shang-Ming Zhou at the University of Plymouth’s Centre for Health Technology, analyzed the records of nearly 70,000 patients collected between 1995 and 2020.
Their custom model, PRECISE-AGZ, initially examined 48,261 possible indicators (from smoking and exercise habits to medication use). It then identified 38 features that together provided the strongest signals of bladder cancer risk.
The system detected 85% of patients who had bladder cancer and correctly classified 91% of those who were cancer-free. It was also more likely to identify the disease than existing NHS referral guidelines.
Why Current Screening Misses Cases
Current referral decisions depend heavily on visible blood in the urine (hematuria). Because this symptom is not unique to cancer, it can send people without the disease for invasive testing while failing to capture others whose medical records contain less obvious warning patterns.
A definitive diagnosis generally requires cystoscopy (where a long, thin tube with a small camera inside is moved up the urethra and into the bladder). More precise risk assessment could therefore help clinicians determine who needs the procedure most urgently and who may be monitored safely.
Rather than relying on one symptom, PRECISE-AGZ considers how multiple pieces of health information interact. The findings suggest that the meaning of a familiar warning sign can change when it is viewed alongside the rest of a patient’s medical history.
Unexpected Bladder Cancer Patterns
The model confirmed established risk factors such as smoking and blood in the urine, but it also uncovered associations that conventional referral systems would be unlikely to detect. People with Parkinson’s disease or dementia showed a lower risk of bladder cancer, while those who used tamoxifen (a breast cancer medication) over long periods showed a higher risk.
These patterns do not establish that the conditions or medication prevents or causes bladder cancer. However, they raise questions about possible biological connections that could be explored in future studies.
Blood in the urine also did not carry the same implications for every patient. Among men with benign prostate enlargement, for example, its presence was associated with a lower cancer risk. If further research confirms such relationships, clinicians may be able to avoid some unnecessary referrals without overlooking people at greater risk.
Three Levels of Bladder Cancer Risk
PRECISE-AGZ placed patients into three categories: low risk (below 7% probability), uncertain (grey zone: 7-55%), and high risk (above 55%). This structure could help health care providers prioritize testing and other resources for patients most likely to have the disease.
People in the uncertain range could be monitored before undergoing clinical intervention, potentially reducing unnecessary cystoscopies. Further investigation and development will be needed before the system can be introduced into routine care, but it could eventually give clinicians a more personalized way to decide who requires urgent testing.
“Bladder cancer ranks as the ninth most common cancer worldwide, with 614,000 new cases diagnosed in 2022. Despite its prevalence, no routine screening program exists for the general population, and—following symptoms—detection relies heavily on invasive cystoscopy procedures.
“Our system goes far beyond traditional symptoms, and it’s great to have uncovered the results that we have. The research shows that the combination of multiple factors—from medication histories to seemingly unrelated conditions—can reveal patterns invisible to conventional screening methods,” said PhD student Xu Wang, who led the data analysis.
Promise and Limits of AI Screening
“Bladder cancer has a major unmet need in early detection, and the prospect of a screening tool is hugely exciting for patients and their families. It could identify those most at risk more accurately, reduce harm from invasive procedures, and catch cancer earlier, when survival outcomes and quality of life can be dramatically improved,” said Dr. Helen Winter, clinical director for the Somerset, Wiltshire, Avon, and Gloucestershire (SWAG) Cancer Alliance.
“This work represents a paradigm shift toward precision screening for bladder cancer. By harnessing the power of interpretable machine learning and comprehensive health records, we’re moving closer to detecting this disease at its earliest, most treatable stages.
“It’s important to emphasize that further validation across different health care systems is essential before anything is rolled out more widely—for example, all of the data analyzed came from the SAIL database in Wales, so we’d want to investigate other data sets, too. But it’s a very exciting early study. Additional studies would also be needed to confirm the causal relationships suggested by the model’s predictions,” said Zhou.
DOI: 10.1109/TBME.2026.3658230
Never miss a breakthrough: Join the SciTechDaily newsletter.
Follow us on Google and Google News.