AKI Mortality Prediction Research
A comprehensive retrospective cohort study applying four machine learning models to predict in-hospital mortality risk in ICU patients with Acute Kidney Injury.
AKI is a sudden reduction in kidney function over hours to days — one of the most frequent and severe complications in ICU patients.
AKI significantly increases ICU mortality. Early risk stratification enables timely intervention that can be life-saving.
Machine learning models combine laboratory and demographic data to estimate individual mortality risk, supporting clinical decision-making.
Raw retrospective ICU data loaded directly from a local PostgreSQL instance. The table contains 2,230 patient records with 30+ raw features including demographics, vital signs, and laboratory measurements.
Exploratory data analysis to verify target balance (~72% survived, ~28% deceased), identify missing value patterns, and detect anomalies. ID columns (row_id, patient_id, protocol_no) are automatically removed.
Categorical variables one-hot encoded. Non-predictive identifier columns removed. Feature matrix prepared with the deathflag target variable extracted separately.
Stratified train-test split preserves the class ratio in both sets. Fixed random seed (42) ensures full reproducibility. No data from the test set ever touches preprocessing fitting.
Multiple Imputation by Chained Equations via the miceforest library. The imputer is fit exclusively on training data (5 iterations) and then applied to both train and test sets — preventing data leakage.
StandardScaler normalizes features to zero mean and unit variance. Fit on train set only, then applied to test set. Required for Logistic Regression and ANN; tree models use unscaled features.
Four models trained in parallel. Evaluated on held-out test set with AUC, Accuracy, F1, Precision, Recall, and MCC. Best model selected by AUC for the prediction interface.
This system was developed exclusively for academic research purposes. It must not be used for clinical decision-making, patient triage, or any medical diagnosis. All model outputs are statistical estimates based on retrospective data and must be reviewed by a qualified clinician before any action is taken. The authors accept no liability for clinical outcomes.