Simple Summary
Renal cancer ranks 14th amongst most fatal cancer diseases listed by WHO worldwide. Apparently, pre-diagnosis of renal cancer plays a role in accurate treatment and a plan of treatment to save the patient can be suggested. The study aims to access the patient information from three different datasets to analyze the prediction of disease using clinical information. The feature engineering techniques will filter the dataset to standardize the prediction process, whereas XAI will identify which features from all the datasets are essential in prediction. The Multilevel Perceptron proves best amongst the algorithms in the prediction of renal cancer with highest accuracy. SHapley Additive exPlanations (SHAPs) visualize the comparison of the features and its impact.
Abstract
Renal cancer is a fatal disease and ranks among the 14 common cancers in the latest WHO ranking. Background/Objectives Renal cell carcinoma (RCC) prediction not only diagnoses the disease at early stages but also prevents patients from suffering pain and eventual death. This study aimed to construct and validate a model that predicts renal cancer using machine learning (ML) classifiers. Methods Feature engineering techniques were used to refine the dataset, such as interaction feature generation, which categorizes the features according to values such as BMI, eGFR, and radiographic size. Binary clinical risk flags identify clinical risks according to the feature values, and correlation pruning removes the common and correlated features among various features. Three public datasets were collected, and the performance of seven machine learning classifiers was evaluated using the proposed model. The evaluation parameters included accuracy, precision, recall/sensitivity, specificity, F1 score, and Matthews Correlation Coefficient (MCC). An Explainable AI method was used to visualize the importance of the model characteristics. Results Among all the algorithms, the Multilevel Perceptron (MLP) algorithm fits best for the prediction of renal cancer with 98%, 96%, and 100% accuracy, and SHapley Additive exPlanations (SHAPs) are used to visualize the importance of all model features. Conclusions The proposed model provides an efficient method and shorter computation time to predict renal cancer using clinical features.