Predicting Sepsis in ICU Patients Using GRU-D and LightGBM for Clinical Decision Support

Sepsis is a life-threatening and acute condition, which is caused by dysregulation in host responses to infection and which can frequently result in organ failure and a high mortality rate in intensive care units (ICUs). Sepsis should be identified early enough to allow timely interventions; however, the conventional scoring system can be limited in accuracy and in its capacity to deal with complex temporal clinical data. This paper proposes a hybrid artificial intelligence system comprising of a gated recurrent unit with decay (GRU-D) and light gradient boosting machine (LightGBM) to predict sepsis in ICU patients. GRU-D model is developed to model the temporal dynamics of the multivariate time-series clinical-time data and can well take into consideration irregular sampling and missing data, whereas LightGBM can use managed features and non-linear interactions to advance predictive performance. The ICU electronic health record data was carefully preprocessed to keep the model reliable and to avoid any information leakage. The hybrid model showed better predictive performance than the standalone GRU-D or LightGBM models, and a better area under the precision-recall curve, higher sensitivity, and earlier sepsis onset. The results show that the hybrid GRU-D + LightGBM model outperforms the individual GRU-D and LightGBM models, achieving the highest accuracy (0.888 ± 0.011), precision (0.857 ± 0.017), recall (0.845 ± 0.020), F1-score (0.851 ± 0.018), and Area Under the Receiver Operating Characteristic Curve (AUROC) (0.931 ± 0.009). Besides, feature importance and SHAP value analyses through interpretability yielded insights into clinically significant predictors of the model, which increases transparency of the model to healthcare providers. The suggested GRU-D and LightGBM model can provide a convenient and effective clinical decision support system that could help clinicians identify and treat sepsis in time and potentially enhance patient outcomes and decrease ICU mortality.