Development and internal validation of a nomogram for in-hospital mortality in older adults with severe trauma: a secondary analysis of a single-center cohort

Mortality after severe trauma remains high in older adults. Models developed in broader trauma populations may not perform similarly in this group. We developed and internally validated a penalized nomogram and assessed calibration, potential net benefit, missing-data uncertainty, model stability, and nonlinearity. We conducted a secondary analysis of the publicly available single-center Peterer et al. cohort. Patients were eligible if they were aged > = 65 years, had Injury Severity Score > = 16, and had a recorded in-hospital mortality outcome. Missing predictors were handled with 20 chained-equation imputations; mortality was used as an auxiliary predictor but was not imputed. The primary ridge model included age, New Injury Severity Score, Glasgow Coma Scale, and lactate. A seven-predictor model was examined as a sensitivity analysis. We used 1000 bootstrap resamples for internal validation and also assessed calibration, nonlinearity, and decision curves. Of 311 source records, 263 met the eligibility criteria and 153 patients died (58.2%). Model A had an apparent area under the receiver operating characteristic curve (AUC) of 0.871 and an optimism-corrected AUC of 0.863 (bootstrap test-AUC interval 0.855–0.873). Its calibration intercept was − 0.006, slope 1.053, observed-to-expected ratio 1.00, and Brier score 0.145; the mean bootstrap calibration slope was 1.006. The restricted cubic spline sensitivity model performed similarly (AUC 0.872; Brier 0.145). Model B had an apparent AUC of 0.888. In complete-score cases, recalibrated Trauma and Injury Severity Score, its age-extended form, and Acute Physiology and Chronic Health Evaluation II had AUCs of 0.802, 0.822, and 0.722, respectively. In this selected single-center cohort, the ridge nomogram showed good discrimination and acceptable overall calibration after internal validation. Its transportability and clinical utility require external validation and prospective impact assessment. Because the source dataset lacked frailty, comorbidity burden, medication use, baseline functional status, and cognitive status, the model should be regarded as a prediction model for older adults with severe trauma rather than a geriatric-specific model.