PREP: a practical RNN-based estimated glomerular filtration rate prediction for inpatients with heart failure

Cardiovascular diseases (CVDs), including heart failure (HF), are the leading cause of mortality worldwide, with an increasing incidence. Chronic comorbidities often accompany CVDs, such as hypertension, diabetes, and renal disease (RD). We propose an application that could assist with the real-time risk detection of renal dysfunction in hospitalized patients with pre-existing heart failure by developing a recurrent neural network (RNN)-based estimated glomerular filtration rate (eGFR) prediction (PREP) model using sequential deep learning (DL) and machine learning (ML) algorithms. We developed a DL-based sequentially predictive model to improve the early identification of RD in patients with HF through creatinine (Cr), the most frequently performed blood test. Additionally, we employed an ensemble model to compress model and improve its performance by reducing weights. We evaluated our model with several metrics including mean absolute error and accuracy. Finally, we proposed the implementation of the Cr-based PREP in web-browser-based applications. The four modules constituting the PREP received the same 24-hour data as input and developed predictions after 12, 24, 36, and 48 hours. PREP accomplished 5.61 of all features to 3.21 of 10 features for MAE of module 1 in the case of regression (eGFR, unit: mL /min/ \\(1.73\\,\\textrm{m}^2\\) ) and 0.82 to 0.89 of Acc for the same module classification. We proposed an ML-based sequentially predictive model and application which are lightweight and practical in assisting with the real-time risk early identification of RD and CVDs during hospitalization. Our findings could be integrated into the information system of hospitals and institutions to improve the management of CVDs. Retrospectively registered. Not applicable.