A Deep Learning Program to Predict Acute Kidney Injury.

Researchers

Journal

Modalities

Models

Abstract

Acute kidney injury is a dangerous and sometime fatal clinical situation, which can cause irreversible damage. If we can predict it earlier and make appropriate prevention before its outbreak, kidney injury could be avoided. One challenge of early recognition of AKI is that the most e-alerts have focused on creatinine-based algorithms, but the elevation of serum creatinine lags behind renal injury. We use recurrent neural network (RNN) to make data mining on laboratory results of MIMIC-III Database. At first, we transfer the case data into Pandas DataFrame of series framed for supervised learning. Then we can use RNN predicts the next serum creatinine values (SCr) based on the last laboratory test results after emergency admissions. We train the RNN on whole dataset (i.e. multi-cases prediction) with LSTM. As the result shown, this prototype can predict criteria (SCr) of AKI with a RMSE (Root Mean Square Error) of 0.017mg/dL.