Hospital Alemão Oswaldo Cruz

Regulatory agencies are increasingly incorporating real-world data (RWD) and real-world evidence (RWE) into decision-making frameworks to complement randomized clinical trials. While some regions, such as the U.S. and EU, have developed structured approaches for RWE use, Brazil's regulatory environment remains comparatively limited. This study examines the status of RWE regulatory integration in Brazil through an analysis of normative documents, institutional publications, and selected case studies, using a comparative policy perspective. Although advances have been made in data standardization and the publication of technical guidelines, such as ANVISA's Guidance No. 64/2023, the practical use of RWE in regulatory processes is still nascent. Key challenges include fragmented data infrastructure, and limited intersectoral coordination. Addressing these gaps will require improved interoperability across health information systems, convergence of methodological standards, and sustained collaboration among regulatory authorities, academia, and data holders to enable consistent and scientifically robust use of RWE in the Brazilian context and, potentially, in other low- and middle-income countries.

Patients with muscle-invasive bladder cancer have varied outcomes after cystectomy. Circulating tumor DNA (ctDNA)-based detection of molecular residual disease may identify patients at high risk for recurrence after cystectomy who can benefit from adjuvant immunotherapy, thus sparing patients at lower risk from unnecessary treatment burden.

In a phase 3, double-blind, randomized trial, we used serial ctDNA testing to monitor (for up to 1 year) patients with muscle-invasive bladder cancer and no radiographic evidence of disease after surgery. Eligible patients who tested ctDNA-positive during surveillance were randomly assigned in a 2:1 ratio to receive intravenous atezolizumab or placebo every 4 weeks for up to 1 year. The primary end point was investigator-assessed disease-free survival. Overall survival was a secondary end point that was assessed in a hierarchical fashion to control for alpha. Patients who persistently tested ctDNA-negative did not receive atezolizumab or placebo.

A total of 761 patients were enrolled; 250 eligible patients who tested ctDNA-positive underwent randomization (167 to the atezolizumab group and 83 to the placebo group). The median disease-free survival was 9.9 months with atezolizumab, as compared with 4.8 months with placebo (hazard ratio for first event of disease recurrence or death, 0.64; 95% confidence interval [CI], 0.47 to 0.87; P = 0.005). The median overall survival was 32.8 months with atezolizumab, as compared with 21.1 months with placebo (hazard ratio for death, 0.59; 95% CI, 0.39 to 0.90; P = 0.01). A total of 28% of the patients who received atezolizumab and 22% of those who received placebo had adverse events of grade 3 or 4 (related to atezolizumab or placebo in 7% vs. 4%); 3% and 2% of the patients, respectively, had fatal adverse events (related to atezolizumab or placebo in 2% vs. none). Among 357 patients with persistent ctDNA-negative status, disease-free survival was 95% at the end of the 1-year monitoring period and 88% at 2 years.

Among patients with muscle-invasive bladder cancer, ctDNA-guided adjuvant therapy with atezolizumab led to significantly longer disease-free survival and overall survival than placebo. (Funded by F. Hoffmann-La Roche; IMvigor011 ClinicalTrials.gov number, NCT04660344.).

To explore whether large language models can assist in refining complex medical definitions, this proof-of-concept study evaluated the capacity of three advanced systems, ChatGPT-4, Gemini 1.5 Pro, and Grok 3, to interpret, critique, and build upon existing definitions of clinical obesity, with particular reference to the Lancet Commission on the Definition and Diagnosis of Clinical Obesity.

Five structured questions were used in a three-stage qualitative design: (1) baseline responses without external references, (2) context-informed responses incorporating the Lancet Commission report, and (3) generative redefinition of clinical obesity. Model outputs were reviewed by a medical expert for accuracy, originality, and concordance with established guidelines.

Without contextual guidance, large language models produced generalized definitions largely reflecting public-domain sources. Supplying the Lancet Commission report yielded more precise, conceptually consistent, and clinically relevant outputs. In the generative phase, all models proposed refined definitions that addressed recognized gaps in existing frameworks, emphasizing functional impairment, psychosocial dimensions, and the limitations of body mass index as a diagnostic criterion.

When anchored to authoritative, domain-specific evidence, large language models can serve as adjunctive tools for knowledge synthesis, clarifying and strengthening complex medical constructs. Integrating human expertise with artificial intelligence under transparent and ethically governed conditions may enable the development of more inclusive, adaptive, and evidence-based clinical definitions.