Deep learning turns scattered bladder cancer notes into measurable BCG outcomes
A natural language processing model built on electronic health records has shown that routine clinical text can be converted into usable outcome data for patients with non-muscle invasive bladder cancer receiving intravesical Bacillus Calmette-Guérin therapy. The effort addresses a persistent problem in bladder cancer care: recurrence and progression are often buried inside narrative notes, procedure reports, and follow-up documentation that are expensive to abstract by hand. For investigators trying to compare treatment effectiveness across real-world cohorts, the bottleneck is not lack of data. It is the shape of the data.
The model used a Bidirectional Encoder Representations from Transformers architecture to classify outcomes drawn from a retrospective Japanese cohort of adults who started BCG induction therapy between April 2016 and June 2022. Human review was incorporated for past history records, creating a hybrid approach in which the model handled large-scale text extraction while clinicians provided targeted support where prior context was essential. The result is a workflow built for clinical reality: tedious enough to require automation, but nuanced enough to still need medical oversight.
Completion of BCG therapy separated recurrence risk from incomplete treatment
The cohort included 372 patients, with 79.3% in the completion group and 20.7% in the non-completion group. The comparison matters because BCG induction is not simply a medication exposure; it is a treatment course whose effectiveness depends on whether patients actually receive the intended sequence. In intravesical therapy, stopping early may reflect toxicity, frailty, logistic barriers, or early clinical deterioration, any of which can distort outcomes if they are not captured carefully in the record.
A single endpoint stood out clearly. Patients who completed BCG had a lower hazard of recurrence than those who did not complete therapy, with a hazard ratio for time to recurrence of 0.40 by multivariate Cox proportional hazards modeling and 0.41 by inverse probability of treatment weighting. Both estimates point in the same direction and suggest a substantial recurrence advantage associated with treatment completion. In a disease where local recurrence drives repeated procedures, surveillance burden, and patient anxiety, that difference has immediate clinical meaning.
BCG acts through local immune activation inside the bladder, and incomplete delivery likely weakens that immune priming. The bladder mucosa needs repeated exposure for the inflammatory cascade to translate into durable antitumor activity, especially in a setting where microscopic residual disease may persist after resection. When the course is interrupted, the biologic pressure against recurrence appears to fall with it.
Human review sharpened the model’s ability to detect progression
Performance improved when human support was added. Before human support, the final BERT model achieved average F1 scores of 0.91 for time to recurrence and 0.74 for time to progression. After human support, those values rose to 0.98 and 0.94, respectively. Progression was the harder endpoint to classify, which is unsurprising in bladder cancer follow-up, where progression events are less frequent and may be documented less consistently than recurrence or surveillance findings.
The need for support in only 10% of documents is an important operational detail. It suggests that the system did not require constant human correction to function at a useful level. Instead, it concentrated expert input where older history or ambiguous documentation could otherwise confuse the model. In practical terms, that kind of selective oversight is the difference between a system that merely impresses in a laboratory setting and one that could scale within a busy cancer center.
The distinction between recurrence and progression also matters clinically. Recurrence captures disease coming back within the bladder, while progression signals a more dangerous shift in biology and prognosis. A model that can separate those endpoints from unstructured text gives researchers a cleaner way to study which patients benefit from BCG, where failures cluster, and how follow-up patterns influence apparent outcomes.
Why this approach matters for real-world bladder cancer research
Retrospective outcome studies in non-muscle invasive bladder cancer have long depended on manual chart review. That process is slow, expensive, and vulnerable to inconsistency when dozens or hundreds of records must be sorted for recurrence timing, progression timing, and treatment completion status. The current model shows that a transformer-based language system can reduce that burden while preserving clinically meaningful signal. More importantly, it can do so using data already embedded in routine care.
This has implications beyond one institution or one therapy. If EHR text can be mined reliably for timing endpoints, investigators can compare treatment strategies at scale without waiting for fully structured registries to be built. In bladder cancer, where surveillance intervals, operative findings, pathology reports, and adjuvant treatment documentation all sit in different parts of the chart, that capability could speed observational research and support quality improvement work.
The model also reinforces a basic oncologic reality: therapy exposure is not binary in practice. Completion and non-completion may reflect a complex mix of tolerability, adherence, and disease behavior, and those factors can shape outcomes as much as the treatment itself. A computational tool that respects those nuances can improve how clinicians interpret real-world effectiveness data.
Limitations remain tied to retrospective text extraction
Even with strong F1 performance and concordant hazard estimates, this remains a retrospective study built around one health system’s documentation patterns and one language environment. The model’s success depends on the quality and structure of the underlying notes, and performance may shift if applied to different institutions, different phrasing habits, or different EHR architectures. The study also required human support for a subset of documents, which means full automation was not achieved.
The main practical message is narrower and stronger than generic enthusiasm about artificial intelligence. A deep learning model can extract recurrence and progression outcomes from bladder cancer records well enough to compare BCG completion with non-completion, and it can do so with limited human assistance. The unresolved questions are deployment and portability: whether the same performance holds outside this setting, and how much local retraining would be needed before the system could be used routinely in other centers.