Active surveillance as the treatment of choice for low-risk prostate cancer: Reliability of results obtained through clinical language processing systems and big data.

INTRODUCTION: Advances in natural language processing (NLP) technologies have gained prominence for extracting relevant clinical information. Savana is a platform capable of analyzing free-text data and interpreting the content of electronic health records (EHRs). OBJECTIVE: To validate the results obtained through NLP by Savana from data of patients with prostate cancer (PC) included in active surveillance (AS), compare them with our database, and assess their reliability. METHODS: Observational and retrospective study of patients with PC in AS between 2014 and 2022. The results from our database were blinded to Savana. Information from the EHRs was transformed by Savana into analysis-ready data. After an initial evaluation, it was necessary to refine the preliminary results and readjust the variables and terminology to eliminate discrepancies. RESULTS: Of the 2865 patients included in our database, 306 met the selection criteria. Savana detected 366 patients with the terms "PC," "Gleason," and "AS." The results were similar regarding Gleason score at diagnosis: 93.4% Gleason 6 in our series vs. 92% in Savana. Likewise, the proportion of patients who received treatment with curative intent, and the type of treatment were comparable: 33.3% in our series (RP: 56.9%; RT: 42.1%) vs. 32.5% in Savana (RP: 59.7%; RT: 40.3%). However, only 24.8% showed Gleason progression in our series vs. 31% in Savana. The mortality rate was 3.2% in our series vs. 7.4% in Savana. CONCLUSIONS: NLP represents a promising tool in clinical research, but its implementation should be approached with caution.