EVALUATING AI-DRIVEN CLINICAL DECISION SUPPORT SYSTEMS FOR EARLY DETECTION OF ACUTE KIDNEY INJURY

Abstract

Acute Kidney Injury (AKI) is a common and severe complication in critically ill patients, with early detection being essential for improving clinical outcomes. This study evaluates the performance, interpretability, and clinical integration of AI-driven Clinical Decision Support Systems (CDSS) for early AKI detection, using a comprehensive electronic health records dataset from a tertiary care ICU. Various machine learning models, including Logistic Regression, Random Forest, Gradient Boosting, and deep learning architectures such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), were assessed. Among these, CNN consistently demonstrated superior performance with an AUROC of 0.89, sensitivity and specificity of 0.86, and an F1-score of 0.85, outperforming all baseline models. Model calibration was robust, with a low Brier score (0.13) and a near-ideal calibration slope (0.98), affirming prediction reliability. The analysis of SHAP values suggested that serum creatinine, BUN and blood pressure played the biggest role which improved the confidence of the model and its usefulness for doctors.  Tests on different datasets confirmed that the CNN’s performance decreases, but only slightly (AUROC of 0.86) compared to the main ones.  Testing the system with ICU clinicians revealed that they were able to respond quickly (in under 12 seconds) and accepted more than 89 percent of the alarms generated.  As a result, AI-CDSS based on CNNs can improve how clinicians decide, take quick actions and minimize expected problems linked to AKI.  For a deployment to be successful, it must first address interpretability, data quality, ethical aspects and should regularly be validated as it is put into clinical use.