Performance of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) in Detecting Acute Kidney Injury (AKI) Among Septic Patients

Sepsis-associated acute kidney injury (SA-AKI) pathophysiology remains uncertain and intricate. However, with the help of machine learning is known to have remarkable implications for disease prediction and detection. With the advancement of machine learning, this research investigates the performance of recurrent neural networks (RNN) and Long Short-Term Memory (LSTM) on 274 adult septic patients with acute kidney injury and without acute kidney injury from Hospital Universiti Sains Malaysia (HUSM), Kelantan. This study used the current definition, depending on Sepsis-3 definition, Kidney Disease: Improving Global Outcomes (KDIGO), Cockroft-Gault, and Intensive Control Insulin Nutrition Glucose (ICING) to train AKI among septic patients based on RNN and LSTM. The results indicated that RNN was effective in identifying AKI in septic patients. The model achieved a high accuracy of 0.7037, a sensitivity rate of 0.7000, a specificity of 0.7059, a precision of 0.5833, and a decent discriminative mean Area Under the Curve (AUC) of 0.6917. According to the result of feature importance, creatinine is the most important feature in defining AKI among septic patients. The performance metrics remained the same regardless of the Glasgow Coma Scale (GCS) being included in the model, although GCS significantly differs between both groups. The performance of the model also deteriorated when <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$S_{I}$</tex> and GCS were removed from the model. In conclusion, although there is a notable statistical distinction in GCS between the two groups, it may not provide any additional predictive value within the RNN model. Regardless of whether GCS is included or not, the ability to compare RNN and LSTM in identifying AKI among septic patients is consistent. Creatinine has been shown to be the most important variable to categorize patients as SA-AKI. <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$S_{I}$</tex>, although shown as having a lower rank, is an important variable to be included in the model. This analysis offers useful insights for future model development, highlighting the importance of feature selection based on machine learning performance rather than only statistical significance.