Delirium modifies the prognostic performance of SAPS-II

We conducted a retrospective cohort study of 10,378 adult ICU patients treated between February 1, 2019 and December 31, 2024. Mortality was predicted using the original SAPS-II equation and the routinely recorded SAPS-II value based on the first 24 h of ICU treatment. The outcome was all-cause mortality during the index hospitalization. Delirium was regularly assessed using the Intensive Care Delirium Screening Checklist (ICDSC). The primary analysis used a post-baseline stratification based on the maximum ICDSC value recorded during the entire ICU stay, complemented by ICD-10 F05 coding: (1) ICDSC unavailable (no valid maximum ICDSC score available), (2) no delirium (MaxICDSC = 0 with complete assessment), (3) subsyndromal delirium (MaxICDSC 1–3 with complete assessment), (4) clinical delirium (MaxICDSC > 3 without F05), and (5) coded delirium (MaxICDSC > 3 with F05). Patients with incomplete assessment and MaxICDSC 0–3 were excluded from the primary analysis and separated into MaxICDSC 0 and MaxICDSC 1–3 groups in a sensitivity analysis. Calibration was summarized using expected/observed (E/O) mortality ratios and LOWESS-smoothed observed-minus-predicted mortality curves. Statistical effect modification was assessed using maximum-likelihood logistic regression with logit-transformed SAPS-II-predicted mortality, delirium-related clinical states, and their interaction; ICDSC unavailable was the reference group. The global interaction was tested by likelihood-ratio comparison. Prespecified robustness analyses used ICDSC classifications restricted to 24, 48, and 72 h, and landmark restrictions requiring an ICU stay of at least 3 days or survival for at least 3 days. Additional calibration and discrimination metrics, risk-band estimates, group characteristics, assessment-opportunity analyses, and an admission-category-adjusted model are reported in the Supplementary Material. Overall, SAPS-II overpredicted mortality (E/O = 2.55; mean predicted risk 37.2% versus observed mortality 14.6%). Miscalibration differed across delirium-related states and predicted-risk strata (Fig. 1 ). Calibration error of SAPS-II across delirium-related clinical states. LOWESS-smoothed observed minus predicted ICU or in-hospital mortality is shown over the supported range of SAPS-II predicted risk. Shaded areas are pointwise 95% percentile-bootstrap confidence bands based on 400 patient-level bootstrap samples; LOWESS span = 0.30 and no additional weighting was applied. The primary phenotype analysis included 8350 patients: ICDSC unavailable, n = 780; no delirium, n = 555; subsyndromal delirium, n = 3384; clinical delirium, n = 1532; and coded delirium, n = 2099. The 2028 patients with incomplete ICDSC assessment and MaxICDSC 0–3 were excluded from this figure At predicted risk < 0.25, mortality was lowest in patients without delirium (3/388, 0.8%), followed by subsyndromal and coded delirium (2.7% each), with marked overprediction (E/O = 13.65, 4.21, and 5.10, respectively). Clinical delirium had higher observed mortality (5.8%) and less overprediction (E/O = 2.14). ICDSC-unavailable patients remained close to calibration in this band (E/O = 0.94). Among patients with predicted risk ≥ 0.75, the point estimate approached calibration in patients without delirium (13 deaths among 18 patients; observed mortality 72.2%; E/O = 1.21, 95% CI 0.71–2.28), whereas persistent overprediction was observed in subsyndromal, clinical, and coded delirium (E/O 2.71–4.82). Because the high-risk no-delirium stratum was small and phenotype assignment depended on subsequent assessment and clinical course, this contrast should not be interpreted as a protective effect of delirium. The interaction between phenotype and logit-transformed predicted risk was significant (likelihood-ratio χ = 62.9, 4 df, p = 7.1 × 10⁻ ). Compared with the ICDSC-unavailable reference group, the risk–mortality association was flatter in subsyndromal, clinical, and coded delirium and steeper in patients without delirium (Table 1 ). Calibration slopes ranged from 1.34 in patients without delirium to 0.38 in coded delirium, while AUC ranged from 0.90 to 0.68, respectively (Supplementary Table S1). The interaction remained detectable with 24-, 48-, and 72-h definitions and in both 3-day landmark analyses; adjustment for admission category did not materially alter it ( χ = 69.9, 4 df, p = 2.4 × 10⁻ ). Separating the 2,028 incompletely assessed patients into MaxICDSC 0 ( n = 850) and MaxICDSC 1–3 ( n = 1178) retained strong evidence of interaction ( χ = 68.3, 6 df, p = 9.0 × 10⁻ ). Incomplete ICDSC 0 was not equivalent to complete no-delirium assessment: mortality was 43.6%, E/O = 1.21, and calibration slope = 0.47, whereas mortality was 10.8% in the group ICDSC 1–3 (E/O = 3.51 and slope = 0.48), (Supplementary Fig. S2 and Tables S9–S11) (Table 2 ).