Early Detection of Sepsis by using Attention based Recurrent Neural Networks with Shapley Additive Exp lanations

In recent years, the severe sepsis and infected shocks are becoming a maj or source of death in Intensive Care Units (ICU), which needs timely diagnosis for the effective treatment outcomes. The medical information of patients is stored in Electronic Medical records (EHR) for improvement of Artificial intelligence (AI) in medicine. However, several existing methods are utilized for sepsis detection including Long Short-Term Memory (LSTM) but mostly suffer with over fitting and interpretability. To overcome these issues, Attention based Recurrent Neural Networks with Shapley Additive Explanations (ARNN-SHAP) is proposed for early detection of sepsis. The SHAP model improves the interpretability of model by higher feature importance using higher covariance of vectors. Initially, the input data is taken from MIMIC III dataset and then preprocessed with different techniques to reduce over fitting. After that, the feature importance is evaluated with the help SHAP method and fed to the ARNN classifier. The ARNN classifier classifies the data into sepsis and non-sepsis labels. From the results, the proposed ARNN-SHAP provided better results when compared to existing CNN-LSTM by providing results in terms of accuracy (90.4<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>), sensitivity (70.6%), specificity (89.7%) and precision (72.8%), Fl-score (70.4%) respectively.