Authors: Assistant Professor P.Rama Krishna, J.Akhila
Abstract: Clinical decision-making, patient flow management, and hospital resource utilisation may all be greatly enhanced with accurate ICU length of stay (LOS) prediction. Healthcare providers may improve operational efficiency, offer prompt treatment planning, and optimise bed allocation with accurate patient stay expectations. This research introduces a machine learning architecture that can be easily explained. Its purpose is to use data taken from EHRs inside CIS to forecast the length of time a patient will spend in the intensive care unit (ICU) when they are admitted. Using actual hospital data, the suggested approach uses supervised machine learning models to divide intensive care unit patients into two groups: those with a short length of stay (LOS) and those with a long LOS. A number of performance metrics are used to evaluate the accuracy, specificity, sensitivity, precision, recall, F1-score, and associated classification measures in order to guarantee a trustworthy model assessment. By differentiating between intensive care unit stays that were brief and those that were lengthy, XGBoost showed the best predictive performance among the machine learning algorithms that were tested, with an area under the curve (AUC) of 98%. In order to make prediction outputs more transparent and interpretable, the framework uses Explainable Artificial Intelligence (XAI). This helps healthcare practitioners understand what elements are impacting the model's predictions. Clinicians, hospital administrators, and resource management teams may rely on the suggested framework's reliable decision-support features, which enhance the functionality of hospital Clinical Information Systems via the combination of accurate prediction and model explainability. In order to promote educated clinical decision-making and efficient healthcare resource planning, the experimental findings show that the suggested method effectively predicts the length of stay in the intensive care unit.
