MACHINE LEARNING APPROACH TO PREDICTING PATIENT OUTCOMES IN INTENSIVE CARE UNITS

Abstract

This study predicts the outcomes of patients in the Intensive Care Units (ICUs) of leading hospitals in Pakistan; this research was focused on the application of various machine learning models on the given data. For this, the study was given a database of 2850 patients in ICU from 5 hospitals of tertiary care for the duration of 2022-2024. The machine learning models, Random Forest, Support Vector Machines, Gradient Boosting, Neural Networks, and Logistic Regression were specifically tailored and tested on ICU mortality, length of stay of the patients in ICU and Mechanical Ventilation usage. The records include a range of age, sex, and vital signs, laboratory data, and clinical intervention within the first 48 hours of ICU admission Clinical endpoints include the ICU Severity Score, measuring and recording clinical intervention and other clinical parameters. The dataset undergoes data cleansing through multiple imputation, feature scaling, the SMOTE algorithm, and other class imbalance techniques. The models were tested through 10-fold cross validation where assessment metrics were precision, recall, score F1, and AUC ROC. The Gradient Boosting model performed best with an accuracy of 89.3 and AUC of 0.91 for mortality prediction. APACHE II score, lactate, and mechanical ventilation were the most feature predictors determined in the feature importance analysis models. To support the clinical decision support systems, clinical SHAP points were also incorporated into the model. The paper outlined the integration of predictive machine learning techniques to forecast ICU outcomes within Pakistan’s healthcare system alongside decision support tools to optimize clinical workflows, resource allocation, and quality of care in the ICU.