Prediction of unplanned 30-day readmission for ICU patients with heart failure

Figures

• 

  Fig. 1. Overview of the methodology. This Figure illustrates the overview of the methodology. The admission, insurance, lab measurements, Elixhauser comorbidity, and the discharge information of the patients were extracted from MIMIC-III database and converted to an event log. The resultant event log was used as an input to the process mining discovery algorithm to produce a process model. The resultant process model along with the event logs were then fed to the DREAM algorithm and resulted in some time information related to the variables. The time information with the demographic and the severity scores of the patients were then fed to a NN to predict unplanned 30-day readmission of the ICU HF patients

• 

  Fig. 2. Architecture of Neural Network (NN). This Figure shows the details of the NN architecture. The timed state samples, demographics information, and the severity scores were fed separately to three branches which each branch contains three hidden layers. A batch normalization layer was added after the first hidden layer of each branch. Also, a dropout with a rate of 20% was used after the first, second, and third hidden layers. At the end, the output layer included softmax activation function to predict unplanned 30-day readmission of ICU HF patients

• 

  Fig. 3. The Mean range of Shapley Values for each variable type. This Figure illustrates the impact of each variable in predicting unplanned 30-day readmission of ICU HF patients. The severity scores (Charlson and Elixhauser) have the highest impact in prediction. Following the severity scores, admission events, demographics, artificial events, comorbidity events and lab measurements events have some impact in prediction in order

• 

  Fig. 4. Ablation study on the variable types

• 

  Fig. 5. Ablation study on the variable types