# Integrating dynamic SOFA changes and age to predict 28-day mortality in ICU patients: a nomogram and machine learning validation study

## TL;DR

This study shows that combining changes in SOFA scores and patient age improves predictions of 28-day mortality in ICU patients, using both a nomogram and machine learning models.

## Contribution

The study introduces a novel nomogram and validates machine learning models that integrate dynamic SOFA changes and age for ICU mortality prediction.

## Key findings

- The nomogram achieved a C-index of 0.852 for predicting mortality in patients with baseline SOFA scores of 4–7.
- The XGBoost model showed an AUC of 0.833 in the internal training set but only 0.671 in external validation.
- ΔSOFA 3–1 was identified as the most influential predictor across datasets via SHAP analysis.

## Abstract

The Sequential Organ Failure Assessment (SOFA) score is widely used to predict prognosis in critically ill patients, but the prognostic value of dynamic SOFA changes (Δ SOFA) and their integration into prediction models remains unclear.

This retrospective study included 665 ICU patients admitted to the Third Hospital of Hebei Medical University between July 2022 and December 2023. The initial and daily SOFA scores (days 1–3) and demographic data were collected. Patients were stratified by SOFA 1 scores (4–7, 8–11, ≥12). A nomogram combining SOFA1, ΔSOFA 3–1, and age was developed, and its discriminative ability and calibration were evaluated. Additionally, an XGBoost model using the same predictors was constructed to explore the potential value of machine learning. External validation was performed using the MIMIC-IV database.

Overall, the 28-day mortality rate was 18.9%. Mortality increased with higher SOFA 1 and ΔSOFA 3–1 scores. The nomogram showed high discriminative ability (C-index: 0.852 for SOFA 1 = 4–7; 0.845 for SOFA 1 = 8–11) and good calibration. The optimized XGBoost model exhibited excellent discriminative performance in the internal training cohort, with an area under the curve (AUC) of 0.833. The AUC was 0.863 in the independent internal test cohort and 0.671 in the external validation cohort. SHAP analysis identified ΔSOFA 3–1 as the most influential predictor across the datasets.

Dynamic changes in SOFA scores (ΔSOFA 3–1), especially in patients with moderate baseline SOFA 1 scores (4–11), significantly improve prognostic accuracy when combined with age. The nomogram provides an intuitive bedside tool for early risk stratification, whereas the XGBoost model demonstrates the potential value of machine learning. External validation highlights the need for further multicenter studies to enhance model generalizability.