The Effect of Missing Data when Predicting Readmission in Heart Failure Patients.

Background: The discharge of patients from hospital care is regulated by guidelines. Still, readmission of heart failure (HF) patients is a common issue, and several calculators have been published to predict it. Aims: We elaborate on how the prediction performance decreases when features become missing. We also elaborate on which features should a user include every time to reach acceptable prediction performance. Method: We prepared a balanced dataset from HF patients in the MIMIC-III database $(N=2,204)$ with 16 features. Using training data (80%) in a four-fold cross-validation manner, we evaluated all feature combinations $(N=Z^{16}-1)$ and found the optimal feature set for the logistic regression model. We also evaluated feature presence in top-performing models (N=655) and identified mandatory features. Finally, we trained the resultant model using all training data and evaluated the effect of missing features ( $N=2^{8}$ combinations) using separate test data (20%). Results: We identified three mandatory features (age, blood urea nitrogen, and systolic blood pressure) and eight optional. This led to a resultant model with eleven features. The hazard ratio (HR) using test data showed a value of 2.08 (95%CI 1.66-2.61) when all eleven features were present. It also showed an HR of 1.73 (95%CI1.39-2.17) when only three mandatory features were present, and others were missing (i.e., replaced by zeros).