Toward a Predictive Model for Medical Residency Entrance Examination Performance in Brazil: A Preregistered Cross-Sectional Study and Exploratory Prototype Score Calculator.

Background In Brazil, admission to medical residency is decided by institution-specific entrance examinations taken at the end of medical school or after graduation, typically after 12-24 months of full-time preparation. Unlike clinical decisions, which are routinely supported by quantitative risk calculators, this preparation receives no quantitative feedback on whether it is on track. We aimed to quantify the association between countable study behaviors and examination performance and, as a proof of concept, to build an exploratory prototype preparation calculator converting those behaviors into a predicted score. Materials and methods Preregistered cross-sectional study of candidates who sat the first phase of the 2026 University of São Paulo Medical School (Faculdade de Medicina da Universidade de São Paulo, FMUSP) residency entrance examination, a 120-item multiple-choice paper held on December 7, 2025. The protocol, instrument, and analysis plan were publicly registered on the Open Science Framework two days before the first response. An anonymous online questionnaire was open from May 13 to May 27, 2026; the outcome was the self-reported percentage score. Spearman correlations were the preregistered primary analysis, with Mann-Whitney U tests and multivariable linear regression. An exploratory prototype was evaluated by leave-one-out cross-validation, with predictor and functional-form selection repeated inside the validation loop. Results Of 53 responses, 49 were analyzed. The mean score was 61.8 ± 6.2%; no respondent scored below 51%, a range containing 31.7% of the candidate population. Four behaviors were associated with score, and all survived false discovery rate correction: practice question volume (Spearman ρ = 0.70, 95% confidence interval (CI) 0.47-0.85, p < 0.001), Anki (Ankitects Pty Ltd, New South Wales, Australia) retention rate (ρ = 0.50, 95% CI 0.22-0.71, p = 0.001, n = 38), months of preparation (ρ = 0.46, 95% CI 0.22-0.65, p < 0.001), and flashcard volume (ρ = 0.43, 95% CI 0.17-0.67, p = 0.002). There was no evidence of an association for video lecture hours (ρ = -0.02, 95% CI -0.31 to 0.27, p = 0.87). An exploratory three-input prototype explained 64.2% of variance (adjusted R²; F(3, 45) = 29.65, p < 0.001) and predicted held-out scores to within 3.2 percentage points on average, against 5.3 for an intercept-only model, with an out-of-sample calibration slope of 0.95 (95% CI 0.72-1.18). Months of preparation showed no evidence of an association independent of retrieval-practice volume (β = 0.12, 95% CI -0.07 to 0.32, p = 0.22). Conclusions Within this selected sample, a small set of countable study behaviors predicted examination performance substantially better than the sample mean, showing that preparation calculators are constructible. The prototype is an exploratory proof of concept, not a validated instrument, and its coefficients are provisional: the design is cross-sectional, baseline knowledge was not measured, and the sample is small, self-reported, and volunteer-based, with no representation of the lowest-scoring third of candidates. It supports no causal or comparative-efficiency claim and should not be applied to candidates outside the observed range. Larger, verified, prospectively sampled data with a baseline assessment are required before such a tool reaches candidates.