Prediction of in-hospital mortality in cardiogenic shock using blood gas analysis data and a neural network model
A Zapolska-Tomasiewicz, A Oleszczak, M Blaziak, W Zimoch, W KuliczkowskiAbstract
Background
Cardiogenic shock is an increasing problem with an in-hospital mortality rate reaching 60%. Early and accurate risk stratification is critical for guiding therapeutic decisions and improving patient outcomes. Arterial blood gas analysis provides rapid and objective information on metabolic and respiratory disturbances in critically ill patients and may support mortality prediction.
Purpose
To develop and validate a neural network–based model using sequential blood gas data to predict in-hospital mortality in patients with cardiogenic shock and enable clinically meaningful risk stratification.
Methods
This retrospective observational study included adult patients with cardiogenic shock admitted to a tertiary academic hospital between January 2021 and December 2024. Demographic and clinical data, including serial arterial blood gas measurements, were extracted from electronic health records. A gated recurrent unit (GRU) neural network was employed to model temporal dependencies in sequential blood gas data and predict in-hospital mortality. The model was trained using the Adam optimizer and feature importance was assessed using permutation feature importance. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC).
Results
A total of 333 patients with cardiogenic shock were included, with an overall in-hospital mortality rate of 56.8%. After exclusion of incomplete datasets, 10,880 arterial blood gas analyses from 298 patients were analyzed. Feature importance analysis identified lactate, base excess, pH, standard base excess, ionized calcium, partial pressure of oxygen, and sodium as the most influential blood gas parameters. Lactate demonstrated the highest relative importance for mortality prediction. The GRU-based model showed high discriminative performance, achieving an AUROC of 0.893.
Conclusions
A GRU-based neural network model using blood gas analysis data can accurately predict in-hospital mortality in patients with cardiogenic shock.