Predicting Readmission of Heart Failure Patients

conference paper

English

Heart failure (HF) is the main reason for readmission in hospitals, especially for elderly patients. To prevent HF recurrence, we propose a method to predict HF probability for patients leaving intensive care units. We use structural data from the freely available MIMIC-III database. We retrieved 2 demographic attributes, 5 physiological measurements from electronic charts, and 10 laboratory features for 7,697 patients. We predict HF with 4 random forest (RF) models at time intervals up to a week, a month, 6 months, and a year. Optimal hyperparameters are calculated for each of the individual models using a grid search on the training set. Next, an ensemble model was constructed from these 4 submodels. The test part of the data (N=1,234) was dichotomized by the ensemble model and survival analysis was performed over a time period of 5.6 years. Results of the log-rank test for dichotomized cohort show a significant difference (p<0.0001) and a Hazard ratio of 3.68 (2.68-5.05). The 4 most important features of the RF model according to the Gini importance namely systolic blood pressure, blood oxygen saturation, blood urea nitrogen, and heart rate are consistent with the parameters observed during discharge of patients from the ICU. Our model also suggests that age and blood glucose play a significant role in predicting HF recurrence.

heart failure, MIMIC, random forest, prediction

KOŠČOVÁ, Z.; VARGOVÁ, E.; PAVLUS, J.; SMÍŠEK, R.; VIŠČOR, I.; BULKOVÁ, V.; PLEŠINGER, F.

26. 12. 2023

IEEE

Atlanta, GA, USA

979-8-3503-8252-5

2023 Computing in Cardiology (CinC)

2325-887X

Computing in Cardiology

United States of America

1

4

4

https://www.cinc.org/archives/2023/pdf/CinC2023-207.pdf