麻豆传媒

In spite of increased understanding about the biology of bladder cancer as well as the development of new treatments, the rate associated with the disease has not changed in some 30 years. In 2021, an estimated 83,730 new cases of bladder cancer were diagnosed in the United States -- 17,200 of which resulted in death.

Bladder cancer mortality increases exponentially with , which carries a 5-year survival rate of 50-60%. About 25% of patients with newly diagnosed bladder cancer have muscle-invasive bladder cancer (MIBC), and another 10-20% have disease progression that eventually leads to muscle invasion.

Pathology reports from transurethral resection of bladder tumor (TURBT) have become critical for the identification of MIBC. Researchers examining bladder cancer treatment patterns and outcomes in note that automated extraction of key information from large databases using machine learning is a much better way to identify muscle invasion compared with manual extraction and individual chart review.

Results from a study assessing a (NLP) model showed that it can quickly and accurately extract key information from big data with an overall accuracy of 92%. When applied to 71,200 patients from the Department of Veterans Affairs (VA) database, the NLP model identified invasion status for 96% of patients with TURBT at the population level. The model also identified 98% of patients with non-MIBC (NMIBC) using clinical and procedural notes as well as pathology reports.

"Considering the very good performance in classifying patients with bladder cancer using a wide range of note types, our NLP model may be a practical and accurate tool for rapid identification of patients on the basis of invasion status, thus aiding in population-based research," Stephen B. Williams, MD, of the Veterans Affairs Health Care System in Durham, North Carolina, and the University of Texas Medical Branch (UTMB) in Galveston, and colleagues wrote in .

Importantly, automated extraction of grade, stage, and quality information from multiple note types means the model can mine a "," the researchers explained. When the NLP model was applied to large datasets across the VA, the model classified 13,642 (19%) patients as having MIBC and 47,595 (66%) as NMIBC. Total processing time was 8 hours and 42 minutes, the team reported.

This 1:3 ratio compares favorably with results from , and "further supports the potential validity of our NLP model," the researchers said, cautioning, however, that the findings may not be generalizable outside of the VA system, and will need to be further validated in other cohorts.

In the following interview, Williams, who is chief of the Division of Urology and director of Urologic Oncology Research at UTMB, elaborated on the findings.

Why has the MIBC mortality rate remained unchanged for so long?

Williams: MIBC heterogeneous disease with biological underpinnings and targeted therapies have only recently come to fruition. Moreover, treatments such as guideline-recommended radical cystectomy or trimodal therapy are offered to only about half of the patients. There is no definitive therapy for those who remain.

Does the use of your NLP model to classify notes from patients who have undergone TURBT have potential for improving clinical practice in any way?

Williams: Although NLP was developed for large population-based research, it may hold promise in the future for clinical management. Identification of muscle invasion is key in patients with bladder cancer, and missing this diagnosis can be lethal.

What is your key message for physicians?

Williams: Given big data and the obvious limitations in assessing it, we need to leverage AI [artificial intelligence], NLP, and machine learning to improve research and inform healthcare delivery.

What's next for your research in this area?

Williams: We will test NLP in other EHR [electronic health record] platforms. Stay tuned -- there's more to come.

Read the study here.