Abstract / Summary
Background/Objectives: We evaluated six-predictor cause-specific models for 45-day in-hospital COVID-19 mortality, including departures from proportional hazards and linearity and transport to Mexican data. Methods: The oxygenation-expanded Barcelona cohort included 1622 adults. Recorded arterial PaO2/FiO2 (P/F) was preferred; otherwise P/F was derived from documented saturation and inspired oxygen using the Rice relationship at SpO2 ≤ 97%. Linear cause-specific Cox models for death and discharge served as a reference. Quadratic and time-varying extensions were assessed by three repetitions of five-fold cross-validation. Frozen Barcelona models were evaluated in a separate Mexican workbook using reconstructed day-45 outcomes and minimum-oxygenation proxies. Historical supplied scores for 634 patients were analysed separately. Results: Barcelona included 325 day-45 deaths. The reference model had apparent area under the receiver operating characteristic curve (AUC) 0.822 and bootstrap-corrected AUC 0.820. Proportional hazards (PH) departures persisted within 45 days. Combining quadratic terms and predictor-by-log(1 + time) interactions improved cross-validated AUC from 0.819 to 0.831 and Brier score from 0.12343 to 0.11974, while calibration slope decreased from 0.945 to 0.830. In 171 Mexican records with 12 reconstructed deaths, AUC was 0.967 for the expanded linear model and 0.857 for the combined extension. Excluding one extreme lactate dehydrogenase (LDH) record yielded 170 patients, 11 deaths, and AUCs of 0.964 and 0.935, respectively. The combined model predicted 17.8% mortality against 6.5% observed in that subset. Conclusions: Flexible effects improved some internal metrics but did not establish an adequate replacement model. Residual misspecification, extrapolation and incompatible predictor timing limit transportability. The calculator implements the linear reference model; definitive admission-time external validation remains necessary.