Kidney Stone Prediction Using Optimized Feature Selection and Multi-Layer Perceptron with Explainable AI

Abstract

Kidney stone disease is a frequent urology condition that occurs from a combination of complex biochemical and physiological factors. This study incorporates a composite decision-support framework combining deep learning and several different feature selection methods to develop an effective kidney stone prediction model. The study used a clinical dataset consisting of 15756 patient records collected from a clinical laboratory. Furthermore, systematic analysis was conducted to determine the most predictive clinical features using three different approaches including: filter-based feature engineering, embedded feature engineering and evolutionary feature engineering. Subsequently, a Multi-Layer Perceptron (MLP) was constructed using only the most predictive features to classify whether there were kidney stones, but also to provide an estimated risk score associated with determining if a patient had kidney stones. SHapley Additive exPlanations (SHAP) was used as a method to enhance model interpretability and clinical trust in the model’s output by providing additional context for making individual patient predictions. Experimental results show that the new Genetic Algorithm feature selection approach for the MLP was able to produce a superior level of classification accuracy, with an overall prediction accuracy of 98.3 percent. Thus, this proposed framework provides a reliable, interpretable, accurate decision-support tool to predict the outcome of kidney stones at the onset of a patient's condition prior to the development of later-stage complications such as acute kidney failure.