Temporal representation of vital signs for mortality prediction after major abdominal surgery: external validation and sampling- resolution analysis

Abstract Background Temporal representation determines which features of routinely recorded vital signs are available for analysis. We evaluated changes in physiological information and mortality prediction associated with different representations after major abdominal surgery. Methods This retrospective study included adults alive and still in the intensive care unit (ICU) 12 hours after admission. Models using first-hour snapshots, 12-hour summaries, or post hoc same-window medians of heart rate, mean arterial pressure, and oxygen saturation were developed in MIMIC-IV and evaluated in the Salzburg Intensive Care database (SICdb) without refitting. A separate within-SICdb analysis compared 5-, 15-, and 60-minute median-based representations in identical densely monitored patients using common cross-validation folds. Processed one-minute records provided the physiological reference. Death was assessed after the landmark through day 28 from ICU admission. Performance differences used patient-paired bootstrap confidence intervals. Results MIMIC-IV included 1740 patients (168 deaths), and SICdb included 2376 (185 deaths). Among 2061 densely monitored SICdb patients (151 deaths), hourly aggregation changed positive to negative abnormality classification in 392/1388 reference-positive patients for hypotension (28.2%), 237/835 for tachycardia (28.4%), and 468/671 for desaturation (69.7%). Paired 5-minute minus 60-minute differences in area under the receiver operating characteristic curve (AUROC) were 0.007 (95% confidence interval − 0.001 to 0.016) for regression and 0.019 (− 0.001 to 0.041) for XGBoost. In external evaluation, full-summary regression achieved AUROC 0.779, with differences of 0.040 (0.019–0.063) versus snapshots and 0.013 (− 0.004 to 0.032) versus same-window medians. External calibration-in-the-large was 0.243 and calibration slope was 1.293. Conclusions Hourly aggregation changed recorded abnormality classifications. AUROC gains from finer representations were not clearly established. Twelve-hour summaries discriminated mortality better than admission snapshots, but their increment over same-window medians was uncertain externally. These findings describe retrospective data representations and do not establish a preferred bedside monitoring frequency.