AI model flags bladder cancer signals up to five years early

Blood in the urine can point to a kidney stone, a prostate problem or bladder cancer. Today, confirming the cancer relies heavily on cystoscopy, in which a camera is moved through the urethra into the bladder. Researchers at the University of Plymouth have now built an AI model that found warning signals in health records as much as five years before official diagnosis—though its effective detection was reported up to 12 months beforehand.

Led by Professor Shang-Ming Zhou at the university’s Center for Health Technology, the team analyzed records from nearly 70,000 patients collected between 1995 and 2020. Its PRECISE-AGZ model sifted through 48,261 potential health indicators, including smoking, exercise and medication use, and selected 38 key features associated with bladder-cancer risk.

The model correctly detected bladder cancer in 85% of patients who had it and correctly identified 91% of cancer-free patients. It also placed people into three groups: low risk, below 7% probability; an uncertain range of 7%–55%; and high risk, above 55%. The researchers say this could help clinicians monitor people in the gray zone before proceeding to cystoscopy, while the model’s findings on Parkinson’s disease, dementia and long-term tamoxifen use point only to associations, not causes.

So what changes in practice? If later studies confirm the results, clinicians could have more information than visible blood in urine alone when deciding who needs urgent investigation. That could help prioritize invasive procedures and identify some cancers earlier, when treatment and quality of life may be better. No routine screening program currently exists for the general population, and the model is not yet a clinical service.

The boundary is clear. All the data came from Wales’s SAIL database, and Zhou says the system needs validation in other health-care systems before wider rollout. The study, published in IEEE Transactions on Biomedical Engineering, is therefore a research model—not a deployed screening tool—but it offers a way to look for risk patterns that conventional referral rules may miss.