P0616REAL-TIME CLINICAL DECISION SUPPORT TO PREDICT AND MANAGE IN-HOSPITAL ACUTE KIDNEY INJURY UTILIZING RECURRENT NEURAL NETWORK

Abstract

Abstract Background and Aims Rapidly Increasing electronic health record (EHR) data and recent development of machine learning methods offers the possibilities of improvement in quality of care in clinical practice. Machine learning can incorporate huge amount of features into the model, and enable non-linear algorithms with great performance. Previously published AKI prediction models have simple design without real-time assessment. Major risk factors in in-hospital AKI include use of various nephrotoxins, repeatedly measured laboratory findings, and vital signs, which are dynamic variables rather than static. Given that recurrent neural network (RNN) is a powerful tool to handle the sequential data, using RNN method in the prediction model is a promising approach. Therefore, in the present study, we proposed a RNN-based prediction model with external validation for in-hospital AKI and aimed to provide a framework to link the developed model with clinical decision supports. Method Study populations were all patients aged ≥ 18 years and hospitalized more than a week at Seoul National University Bundang Hospital (SNUBH) from 2013 to 2017 (training cohort) and at Seoul National University Hospital (SNUH) in 2017 (validation cohort). All demographics, laboratory values, vital signs, and clinical conditions were obtained from the EHR of each hospital. A total of 102 variables included in the model. Each variable falls into two categories: static and dynamic variable; static variable was time-invariant values during hospitalization, and dynamic variables were daily-updated values. Baseline creatinine was determined by searching the minimum serum Cr level within 2 weeks before admission. We developed two different models (model 1 and model 2) using RNN algorithms. The outcome for model 1 was the occurrence of AKI within 7 days from the present. In model 2, we constructed the prediction model of the trajectory of Cr values after 24 hours, 48 hours, and 72 hours, using available Cr values from 7 days ago to the present. Internal validation was performed by 5-fold cross validation using the training set (SNUBH), and then external validation was done using test set (SNUH). Results A total of 40,552 patients in training cohort and 4,000 patients in external validation cohort (test cohort) were included in the study. The mean age of participants was 62.2 years in training cohort and 58.7 years in test cohort. Baseline eGFR was 93.8 ± 40.4 ml/min/1.73m2 in training cohort and 88.4 ± 23.2 ml/min/1.73m2 in test cohort. In model 1 for the prediction of AKI occurrence within 7 days, the area under the curve was 0.93 (sensitivity 0.90, specificity 0.96) in internal validation, and 0.83 (sensitivity 0.83, specificity 0.82) in external validation. The model 2 predicted the creatinine trajectory within 3 days accurately; root mean square error was 0.1 in training cohort and 0.3 in test cohort. To support the clinical decision for AKI manage, we estimated the predicted trajectories of future creatinine levels after renal insult removal, such as nephrotoxic drugs, based on the established model 2. Conclusion We developed and validated a real-time AKI prediction model using RNN algorithms. This model showed high performance and can accurately visualize future creatinine trajectories. In addition, the model can provide the information about modifiable factors in patients with high risk of AKI.