Temporal alignment and dynamic clinical status reshape associations between first-week antibiotic trajectories and in-hospital mortality after cardiopulmonary bypass

First-week antibiotic use after cardiopulmonary bypass (CPB) evolves with postoperative illness, whereas fixed-window classification can overlap outcomes. We examined how temporal alignment and contemporaneous clinical-status adjustment changed trajectory–outcome associations and whether trajectories added information beyond simpler exposure summaries. This retrospective cohort included adults undergoing CPB with hospital stays exceeding 24 h in MIMIC-IV, INSPIRE and SICdb. Eight trajectories spanned [0, 168 h). We compared fixed-window, day-7 landmark, cause-specific death and live-discharge, lagged person-day and source-specific models with clinical-status adjustment. Day-7 corrected discrimination was compared among clinical-indicator, exposure-day-count and trajectory models. Among 14,956 patients, 441 (2.9%) died in hospital. Relative to short-course prophylaxis, fixed-window adjusted odds ratios (aORs) for persistent broad-spectrum exposure, late escalation and early cessation or de-escalation were 9.78 (95% CI 6.43–14.87), 9.20 (6.24–13.55) and 3.46 (1.78–6.71); landmark aORs were 11.29 (6.76–18.86), 9.46 (5.85–15.30) and 1.46 (0.43–4.96). In 5,299 MIMIC-IV/SICdb landmark patients, initial-severity-adjusted cause-specific death hazard ratios for persistent exposure and late escalation were 3.71 (2.07–6.67) and 3.52 (1.99–6.22), decreasing to 1.13 (0.55–2.32) and 1.05 (0.55–2.01) after adjustment for clinical indicators. Live-discharge hazard ratios were 0.66 (0.57–0.77) and 0.82 (0.73–0.92). Across 100,537 day-3-or-later person-days, one- and three-day-lag broad-spectrum exposure aORs were 0.60 (0.36–0.99) and 0.79 (0.47–1.36), with source-dependent estimates. In 1,036 monitored/measured day-7 patients, corrected AUCs were 0.837 for clinical indicators, 0.835 with exposure-day count and 0.832 with trajectories; the trajectory-minus-count difference was − 0.0034 (95% bootstrap interval − 0.0068 to 0.0144). Temporal alignment reduced the early cessation or de-escalation estimate. Contemporaneous clinical adjustment attenuated death associations, and daily and source-specific estimates varied. Trajectories did not demonstrate improved discrimination over clinical indicators or exposure duration. They describe treatment and postoperative course, not an independent antibiotic effect or a validated prognostic tool. Not applicable.