AI identified 85% of bladder cancer cases in Welsh health-record study
PRECISE-AGZ searched decades of medical records for risk patterns. Signals traced back five years, but that is not the same as demonstrating reliable diagnosis five years earlier.
In a retrospective analysis of nearly 70,000 Welsh patients, PRECISE-AGZ correctly classified 85% of bladder cancer cases and 91% of patients without cancer. Researchers reported useful detection as early as 12 months before recorded diagnosis; clues found as far back as five years do not establish equivalent performance throughout that period. The model’s risk tiers could help prioritize cystoscopy, but monitoring some patients and reducing invasive tests remain proposed benefits. Since the study used one Welsh database, external validation is needed before clinical deployment.
01
The team assessed 48,261 possible indicators in records from 1995–2020 and narrowed them to 38 predictive features.
02
PRECISE-AGZ assigns low risk below 7%, an uncertain range of 7–55%, and high risk above 55%.
03
The model linked tamoxifen use with higher risk and Parkinson’s disease and dementia with lower risk; these associations do not establish cause or protection.
Routine medical histories may contain bladder cancer clues long before diagnosis. University of Plymouth researchers developed an AI model that identified 85% of cancer cases in a retrospective study of nearly 70,000 patients in Wales. The findings suggest a way to sharpen referrals for invasive testing, rather than a replacement for the procedure used to confirm cancer.
One symptom versus a medical history
Visible blood in the urine is a major warning sign under current referral guidelines. But kidney stones and prostate conditions can cause it too, complicating decisions about who needs further investigation. Confirming bladder cancer requires cystoscopy: passing a thin tube with a camera through the urethra into the bladder.
The Plymouth team’s model, PRECISE-AGZ, takes a broader view. It searches electronic health records for combinations of symptoms, conditions, risk factors and healthcare activity. Researchers assessed 48,261 possible indicators, including smoking, exercise and medication use, then narrowed them to 38 features useful for predicting bladder cancer risk.
Led by Professor Shang-Ming Zhou, the researchers examined records covering 1995–2020 from the SAIL database in Wales. The study appeared in IEEE Transactions on Biomedical Engineering. In a University of Plymouth account published by ScienceDaily, the team reported better cancer detection than existing NHS referral guidelines.
Two sides of the classification result
85%Patients with bladder cancer
Proportion the model correctly identified as having bladder cancer in the study.
91%Patients without bladder cancer
Proportion the model correctly classified as not having the disease in the study.
A five-year signal is not a five-year diagnosis
The timing needs careful distinction. Researchers reported effective detection as early as 12 months before a patient’s recorded diagnosis. Some warning signals could be traced back as far as five years. Those are different findings: a historical clue does not mean the model demonstrated the same detection performance throughout that longer window.
The model recognized established risk factors, including smoking and blood in the urine. It also found less expected associations: Parkinson’s disease and dementia were linked to lower bladder cancer risk, while long-term use of the breast cancer drug tamoxifen was linked to higher risk. These associations do not show that the conditions protect against cancer or that the drug causes it.
Even a familiar warning sign changed meaning with context. Among men with benign prostate enlargement, blood in the urine was associated with lower bladder cancer risk. The researchers see such relationships as a possible way to distinguish patients needing urgent investigation from those who might avoid invasive testing.
Sorting risk before ordering a procedure
Beyond a yes-or-no classification, PRECISE-AGZ divided patients into three groups based on their estimated probability of having bladder cancer:
Low risk: below 7% probability.
Uncertain, or grey zone: 7–55% probability.
High risk: above 55% probability.
The proposed benefit is more selective testing. Researchers suggest that patients in the uncertain group might receive monitoring before cystoscopy, while those at greatest risk could receive more immediate attention. Reducing unnecessary examinations remains a potential benefit, dependent on further research confirming the system’s reliability—not an established outcome of using it in care.
For now, the model is not ready for widespread clinical use. All the records came from one Welsh database; the researchers say validation across other patient populations and healthcare systems is essential before wider rollout. The next test is whether those risk judgments hold up beyond the setting in which they were developed.
Sources
sciencedaily.comAI uncovers hidden bladder cancer warning signs years before diagnosis
Reader comments
Newest comments first. Replies stay oldest first.