#2886 Machine learning-based assessment of serum creatinine and NT-proBNP levels for predicting mortality in patients with heart failure and kidney injury

Abstract

Abstract Background and Aims In a population hospitalized for acute heart failure, about 20% are affected by acute kidney injury, a significant independent factor linked to increased mortality both during hospitalization and within one year. While existing research offers insights, crafting a precise model to predict mortality in heart failure cases remains a complex task, primarily due to the difficulties in pinpointing critical determinants and achieving high accuracy. The aim of our research was to assess the efficacy of serum creatinine and NT-proBNP levels as predictors for mortality in heart failure patients, using machine learning. Method In our retrospective cross-sectional study, we examined the records of 1,042 patients hospitalized due to acute heart failure at the Institute of Cardiovascular Diseases of Vojvodina, spanning the period from January 1, 2021, to December 31, 2022. The data was initially cleaned, after which we employed a Random Forest model for our analysis. The dataset was partitioned into a training subset, constituting 70% of the data, and a testing subset, comprising the remaining 30%. To achieve optimal precision, we ran the model 100 times. Key performance metrics, including mean sensitivity, specificity, accuracy, F1 score, Matthew's correlation coefficient (MCC), and the area under the Receiver Operating Characteristic (ROC) curve (AUC), were derived from the model to assess its efficacy. Results Following data cleansing, the dataset encompassed a total of 389 patients, of which 58% were male. The median age in this cohort was 71 years, with an interquartile range (IQR) of 63 to 79 years. Key laboratory findings included a median serum creatinine level of 111 µmol/L (IQR: 85-167 µmol/L) and a median NT-proBNP level of 6,589 pg/mL (IQR: 2,822-15,608 pg/mL). After the partition, the median number of patients in the training set was 272 and the testing set comprised 117 patients. The performance of the Random Forest model was characterized by a sensitivity of 0.530, specificity of 0.916, an overall accuracy of 0.856, and an F1 score of 0.516. The MCC stood at 0.439 with the AUC under the ROC curve being 0.713. Conclusion While the relatively low sensitivity observed in our model could be attributed to the generally lower levels of serum creatinine upon admission and the retrospective nature of the study, which may have influenced the model's ability to identify all true positive cases effectively, serum creatinine and NT-proBNP levels hold promise as effective predictors for mortality in heart failure patients.