Risk prediction for heart failure rehospitalization using deep learning with real-world data

초록

Background:Heart failure rehospitalization (HFH) after acute decompensation is closely related to heart failure death. The risk assessment of risk for HFH is important for managing and treating patients with HF. Therefore, some risk prediction models have been introduced. However, there was no study for using deep learning method with real-world data. The aim of study was to develop the prediction model for HFH within 30, 90, and 365 days after acute HF (AHF) discharge. Methods: We analyzed the data of patients admitted due to AHF between January 2014 and January 2019 in an university hospital. To select potential independent features, we used univariate analysis based on statistical tests including t-test,chi-squared and Kolmorogov Smirnov test and then used Recursive Feature Elimination with Cross-Validation. In performing deep learning-Based survival algorithms, we use hyperbolic tangent activation layers followed after recurrent layers with gated recurrent units. The prediction model was trained with the batch size set to 20. To assess the 30-day, 90-day, and 365-day readmission prediction, we used the AUC, precision, recall, specificity, and F1 measure. We applied Shapley value to identify which features contributed to HF readmission and were the most important at various scales or levels. Results:The data of 919 patients aged≥20 years were collected. Two hundred eighty-eight patients were readmitted within 365 days of discharge or follow-up visit date. Twenty-two prognostic features that exhibit statistically significant associations with HFH were identified (age, blood pressure, hospital stay,the previous history of hypertension, diabetes, atrial fibrillation and coronary artery disease, WBC and platelet count, the level of hemoglobin, sodium, albumin, creatinine, use of intravenous furosemide during index admission and the use of angiotensin-converting-enzyme inhibitors or angiotensin receptor blocker, beta-blocker, mineralocorticoid receptor antagonist, diuretics and anticoagulants).When the acceptable percentage for missing data is 30%, the AUC value had shown moderate discrimination (AUC: 0.63, 0.74, and 0.76, respectively, for predicting readmission within 30, 90, and 365 days of follow-up [FU]). The contribution of each feature for HFH within 365-days of FU is shown in Figure by order of importance. The features at the follow-up period have a relatively higher contribution to HFH than features from other time points. Conclusion:Our deep learning-based model using real-world data could provide valid predictions of HFH in 1yr FU. And It can be easily utilized to guide appropriate interventions or care strategies for patients with HF. The closed monitoring and blood test in daily clinic are important and are helpful for accessing the risk of HFH.