Introduction. Chronic pelvic pain syndrome (CPPS) is a heterogeneous multi-domain condition; both the European Association of Urology and the American Urological Association recommend clinical phenotyping along the seven UPOINTS domains: Urinary, Psychosocial, Organ-specific, Infection, Neurologic/systemic, Tenderness of skeletal muscles, and Sexual. Manual phenotyping of large clinical streams is impractical; the application of large language models (LLM) to this task in the Russian-language clinical setting has not been previously described.
Objective. To develop and pilot an automated approach based on a large language model coupled with a specialized urological knowledge base for UPOINTS-based phenotyping of CPPS patients from Russian-language online consultation text; to characterize the observed domain distribution; and to assess the use of the Meares–Stamey test in real ambulatory practice.
Material and Methods. Observational retrospective study on the full sample of 2.733 online consultations addressed to a single Russian urologist (27.08.2025–16.04.2026). The study protocol was prospectively registered. Each consultation passed through five sequential processing stages: local two-layer deidentification, scope classification, classification across 18 nosological categories, generation of a clinical vignette, and UPOINTS phenotyping with retrieval of relevant context from a urological knowledge base. The primary language model was an external commercial large language model (DeepSeek Chat v3.1); all calls to it were made exclusively on deidentified text. A secondary analysis of Meares–Stamey test use was performed via automated pattern search followed by LLM status validation. 95% confidence intervals were computed by the Wilson method.
Results. A CPPS phenotype was identified in 281 patients (10.3%; 9.2–11.5). Positive domain rates: Urinary 50.2%, Sexual 35.9%, Psychosocial 30.6%, Organ-specific 27.4%, Tenderness 26.7%, Infection 21.4%, Neurologic 11.0%. In 74.4% of patients the phenotype was formed by one to three positive domains. Red flags were noted in 10.7%. The full 4-glass Meares–Stamey test was not registered in any of 2,733 consultations; the 2-glass Nickel test in one; expressed prostatic secretion after prostate massage without VB-fraction control in 25 cases. In 70% of prostatitis-related consultations the consultant recommended the test to the patient, while it was actually performed 3–5 times less frequently.
Conclusion. Coupling a large language model with a specialized urological knowledge base permits UPOINTS phenotyping of large textual streams, with clearly defined methodological limitations: a positive Infection domain in text material is not equivalent to laboratory-verified bacterial prostatitis. The rarity of the full Meares–Stamey test in real ambulatory practice is empirically confirmed. Prospective validation against manual expert annotation is required.
