Does Replacing SOFA-1 with SOFA-2 Change Mortality Discrimination and Calibration Across Patient Subgroups? A Paired Analysis in MIMIC-IV
- 1Harvard University
Abstract
BackgroundThe Sequential Organ Failure Assessment (SOFA) score describes organ dysfunction in intensive care. SOFA-2, published in 2025, updates the 1996 score (SOFA-1). It is unknown whether the update changes mortality discrimination and calibration within patient subgroups.
MethodsWe analyzed 85,228 first ICU stays per hospitalization (65,354 adults) in MIMIC-IV v3.1. We scored SOFA-1 (two prespecified conventions for intubated patients) and SOFA-2 from the same data over the same first 24 hours. The primary outcome was in-hospital mortality; ICU mortality was secondary. Using a paired patient-level bootstrap (5,000 replicates), we estimated AUROC overall and by sex, race and insurance; the paired change between scores; subgroup gaps in AUROC and the change in those gaps; and descriptive calibration. The protocol was committed before any analysis relating scores to mortality.
ResultsIn-hospital death occurred in 9,468 stays (11.1%). AUROC was 0.762 for SOFA-2, 0.743 for SOFA-1 as charted and 0.735 for SOFA-1 under the community convention; SOFA-2 minus SOFA-1 was +0.019 (+0.016 to +0.023) and +0.027 (+0.023 to +0.031). Swapping single domains showed that the respiratory domain accounted for the whole gain, about 74% of it through the SpO2/FiO2 fallback. Subgroup gaps changed against SOFA-1 as charted (male-female +0.021; White-Black +0.025), driven by the brain-domain rule, but not against the community convention. The White-Black change was not seen in patients under 65, and neither change was consistent across ICU types. Calibration differences between groups were large but changed little between scores (|{Delta} log O/E| [≤] 0.045). Insurance-group differences largely disappeared in patients under 65, whereas race-group differences did not.
ConclusionsSOFA-2 modestly improved overall discrimination for in-hospital mortality, mainly through its respiratory domain. Changes in subgroup performance gaps were small. They depended on how SOFA-1 handles intubated patients, partly reflected age, and were not consistent across ICU types. These are descriptive associations in a single database; we draw no causal or fairness conclusions.
Related
- Composite Haematological and Nutritional Indices Show No Evidence of Improved Long-Term Mortality Discrimination Beyond Their Own Components in Non-ST-Segment Elevation Myocardial Infarction
- Comparative Discrimination of Boey, PULP, and PPUMSs for Textbook Outcome Failure, Failure-to-Rescue, and 30-Day Mortality Following Perforated Peptic Ulcer Surgery
- Auxiliary modalities add no detectable discrimination or calibration over structured EHR for 30-day COPD readmission: evidence from a leakage-audited MIMIC-IV cohort
- Cross-cultural adaptation and comparative psychometric evaluation of four patient-reported outcome measures for breast cancer-related lymphedema: measurement properties and clinical discrimination
- Paradoxical inflation of prediction model performance measures in minority patient subgroups: the role of upstream bias.
BibTeX
@misc{balakrishna2026does,
title = {{Does Replacing SOFA-1 with SOFA-2 Change Mortality Discrimination and Calibration Across Patient Subgroups? A Paired Analysis in MIMIC-IV}},
author = {K. Balakrishna and V. Bhatt and A. Hammond and R. Pirracchio and A. H. Zai},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.23.26366238},
url = {https://www.medrxiv.org/content/10.64898/2026.10.23.26366238v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Does Replacing SOFA-1 with SOFA-2 Change Mortality Discrimination and Calibration Across Patient Subgroups? A Paired Analysis in MIMIC-IV
AU - K. Balakrishna
AU - V. Bhatt
AU - A. Hammond
AU - R. Pirracchio
AU - A. H. Zai
PY - 2026
DA - 2026/10/23
PB - medRxiv
DO - 10.64898/2026.10.23.26366238
UR - https://www.medrxiv.org/content/10.64898/2026.10.23.26366238v1
AB - BackgroundThe Sequential Organ Failure Assessment (SOFA) score describes organ dysfunction in intensive care. SOFA-2, published in 2025, updates the 1996 score (SOFA-1). It is unknown whether the update changes mortality discrimination and calibration within patient subgroups. MethodsWe analyzed 85,228 first ICU stays per hospitalization (65,354 adults) in MIMIC-IV v3.1. We scored SOFA-1 (two prespecified conventions for intubated patients) and SOFA-2 from the same data over the same first 24 hours. The primary outcome was in-hospital mortality; ICU mortality was secondary. Using a paired patient-level bootstrap (5,000 replicates), we estimated AUROC overall and by sex, race and insurance; the paired change between scores; subgroup gaps in AUROC and the change in those gaps; and descriptive calibration. The protocol was committed before any analysis relating scores to mortality. ResultsIn-hospital death occurred in 9,468 stays (11.1%). AUROC was 0.762 for SOFA-2, 0.743 for SOFA-1 as charted and 0.735 for SOFA-1 under the community convention; SOFA-2 minus SOFA-1 was +0.019 (+0.016 to +0.023) and +0.027 (+0.023 to +0.031). Swapping single domains showed that the respiratory domain accounted for the whole gain, about 74% of it through the SpO2/FiO2 fallback. Subgroup gaps changed against SOFA-1 as charted (male-female +0.021; White-Black +0.025), driven by the brain-domain rule, but not against the community convention. The White-Black change was not seen in patients under 65, and neither change was consistent across ICU types. Calibration differences between groups were large but changed little between scores (|{Delta} log O/E| [≤] 0.045). Insurance-group differences largely disappeared in patients under 65, whereas race-group differences did not. ConclusionsSOFA-2 modestly improved overall discrimination for in-hospital mortality, mainly through its respiratory domain. Changes in subgroup performance gaps were small. They depended on how SOFA-1 handles intubated patients, partly reflected age, and were not consistent across ICU types. These are descriptive associations in a single database; we draw no causal or fairness conclusions.
ER -