AI Tool Accurately Predicts Kidney Injury Signs In Critically Ill Patients

Acute kidney injury (AKI), characterized by a rapid increase in serum creatinine or a decrease in urine output, is a primary cause of complications and increased mortality among patients in the intensive care unit (ICU). Despite the importance of early detection and intervention in AKI, current monitoring methods like vital signs, blood tests, and urine analysis, fall short of offering effective solutions. Serum creatinine, a common diagnostic tool for AKI, is not always reliable for early detection. The rise of artificial intelligence (AI) has led to numerous machine-learning models that have shown high accuracy in predicting outcomes for ICU patients, including AKI detection. However, the use of machine learning to predict oliguria, a critical component of AKI that is associated with higher mortality, has not been extensively researched.

Researchers at Chiba University Graduate School of Medicine (Chiba, Japan) have developed a machine-learning model that could predict the onset of oliguria in ICU patients. They developed the model and assessed its accuracy using data from a large, single-center surgical/medical mixed ICU. The model was based on 28 clinically relevant variables, including urine output, SOFA score, serum creatinine, pO2, FDP, IL-6, and peripheral temperature. It showed a high Area Under the Curve (AUC) of over 0.90 for predicting oliguria between 6 to 72 hours. This high accuracy was attributed to the large dataset of over 10,000 patients, providing extensive training data. The model’s high accuracy and capability to predict oliguria over longer periods with the AUC remaining unchanged even after reducing the variables in the model development indicate its robustness.

Image: The machine-learning model predicts oliguria in critically ill patients (Photo courtesy of 123RF)

In addition, the method of predicting the onset of oliguria from an arbitrary time could have improved the accuracy by increasing the number of training datasets. The model was built based on 28 clinically relevant variables although the overlap of the top-listed variables in the model with those in a dataset of 1,018 values supports the viability of the chosen variables for prediction. Given that oliguria can identify AKI earlier than serum creatinine and is linked to poor outcomes in critically ill patients, this machine-learning model could be instrumental in early AKI detection. This early detection could lead to better patient management and timely interventions, potentially improving the prognosis for this patient group.

Related Links:
Chiba University Graduate School of Medicine