Journal Article

FZJ-2026-02836

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Predicting a sudden decrease in oxygenation in mechanically ventilated adult intensive care patients

Polzin, R. (Corresponding author) ; Fritsch, S. J.FZJ* ; Sharafutdinov, K. ; Bickenbach, J. ; Marx, G. ; Schuppert, A.

2026
IOP Publishing Bristol

This record in other databases:  

Please use a persistent id in citations: doi:10.1088/3049-477X/ae7cfe  doi:10.34734/FZJ-2026-02836

Abstract: Hypoxemia in mechanically ventilated ICU patients can evolve rapidly, and early detection is critical for timely intervention. We developed and internally validated a machine-learning clinical decision support system to predict near-term oxygenation deterioration using retrospective data from a single university hospital ICU (March 2020–September 2022). The cohort included 3,676 adults who received ≥ 24 h of cumulative invasive mechanical ventilation; analyses were performed in 3,267 adults with complete oxygenation data (65.84% male; 62.4% aged <70 years, 24.6% 70–79, 13.0% ≥ 80). Because ARDS onset could not be adjudicated without chest imaging, the primary endpoint was rapid oxygenation loss between two non-overlapping 24-hour windows separated by 48 hours, defined as either (i) an absolute decline in mean daily PaO2/FiO2 >100 mmHg when baseline >450 mmHg or >60 mmHg when baseline >350 mmHg, or (ii) crossing thresholds from >250 to <200 mmHg or from >150 to <100 mmHg. Models were trained with 5-fold hyperparameter tuning and evaluated on a stratified, 10% patient/ICU-visit hold-out test set (event rate 11.2%). The gradient-boosted tree model (XGBoost) achieved ROC AUC=0.89 and PR AUC=0.52, outperforming a baseline logistic regression model. Probabilities were calibrated on the validation split (Brier=0.095; ECE=0.0004). At a 0.10 risk threshold, sensitivity 0.776, specificity 0.844, PPV 0.387, NPV 0.967, and alert rate 22.6%; at 0.20, sensitivity 0.658, specificity 0.911, PPV 0.484, NPV 0.955, and alert rate 15.3% were achieved. By modeling oxygenation dynamics rather than static cutoffs, the system identifies patients at risk for clinically significant hypoxemia up to 72 hours in advance. These results support the feasibility of forecasting oxygenation loss in ventilated ICU patients; external validation and prospective evaluation are needed to assess generalizability and clinical impact.

---

Contributing Institute(s):

1. Jülich Supercomputing Center (JSC)
2. Center for Advanced Simulation and Analytics (CASA)

Research Program(s):

1. 5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs) and Research Groups (POF4-511) (POF4-511)
2. SDI-S - SDI-S: Smart Data Innovation Services - Experimentelle Erprobung und Entwicklung von KI-Dienstverbünden für Innovationen auf industriellen Daten (01IS22095D) (01IS22095D)

Appears in the scientific report 2026

---

Database coverage:
;

---