An interpretable machine learning model for real-time sepsis prediction based on basic physiological indicators

In view of the important role of risk prediction models in the clinical diagnosis and treatment of sepsis, and the limitations of existing models in terms of timeliness and interpretability, we intend to develop a real-time prediction model of sepsis with high timeliness and clinical interpretability.

We used eight real-time basic physiological monitoring indicators of patients, including heart rate, respiratory rate, oxygen saturation, mean arterial pressure, systolic blood pressure, diastolic blood pressure, temperature and blood glucose, extracted three-hour dynamic feature sequences, and calculated 3 linear parameters (mean, standard deviation, and endpoint value), a 24-dimensional feature vector was constructed, and finally a real-time sepsis prediction model was constructed based on the Local Interpretable Model-Agnostic Explanation (LIME) interpretability method.

The area under the receiver operating characteristic curve (AUROC), Accuracy and F1 scores of Extremely Randomized Trees we built were higher than those of other models, with AUROC above 0.76, showing the best performance. The Imbalance XGBoost has a high specificity (0.86) in predicting sepsis. The LIME local interpretable model we built can display a large amount of valid model prediction details for clinical workers' reference, including the prediction probability and the influence of each feature on the prediction result, thus effectively assisting the work of clinical workers and improving diagnostic efficiency.

This model can provide real-time dynamic early warning of sepsis for critically ill patients under supervision and provide a reference for clinical decision support. At the same time, interpretive analysis of sepsis prediction models can improve the credibility of the models.