NefroAssist: A Machine-Learning Prediction Model for Early Nephrology Intervention in Hospitalized Patients at Christus Muguerza Group

Background: Delay in nephrology consultation is associated to severe AKI stages and critical illness, resulting in urgent RRT, increased mortality, reduced renal recovery, greater dialysis dependence, and higher costs. To prevent AKI, based on the article “Acute Kidney Injury Risk Assessment and the Nephrology Rapid Response Team”, which describes a model based in the “Fantastic 4”, to identify patients at high risk for AKI; we developed NefroAssist. This machine learning algorithm has the objective to predict the likelihood of requiring nephrology consultation for patients within the first 6 hours of admission. It employs information from the electronic medical record and allows collaboration with a nephrologist to improve outcomes. Methods: A predictive model was developed with database from the electronic medical record system available from 2023 of 4 hospitals. The methodology for handling the data was a ETL process (extract, transformation, and load). The initial load data were 23,104 individual records. Through the elimination of irrelevant data, data imputation, and the creation of new data, a split of 80-10-10 was performed. Where 80% was allocated for training, 10% for data validation, and 10% for testing the models. In the evaluation, the logistic regression model with has been able to fit correctly the data. Results: Using logistic regression we train a model which demonstrated optimal performance in classifying patients who will require a nephrology consultation. With a specificity of 89% and sensitivity of 73%. The factors identified from the F4 were F1: polytrauma, antibiotics, coronary angiography, major surgery, and sepsis; F2 diabetic and hypertensive patients; and F4 creatinine, proteinuria. Conclusion: Nefroassist successfully predicted which patients will require nephrology intervention, with a sensitivity of 73% and a specificity of 88%. We have validated it retrospectively, the next step is to apply and validate it prospectively.