Inductive Multivariate Convolution Neural Network Towards Early Diagnosis and Classification of Chronic Kidney Disease

Chronic Kidney Disease (CKD) is a lifethreatening condition caused by impaired kidney function or renal disease. According to the World Health Organization, CKD affects <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10{\text{\% }}$</tex> of the global population and results in high mortality due to late diagnosis. Although many machines learning and deep learning models have been proposed to detect and classify CKD, they often struggle with transforming input features to the Glomerular Filtration Rate (GFR) and handling non-linear feature embedding for accurate predictions. This research introduces a new methodology for early CKD diagnosis and classification using a Multivariate Convolutional Neural Network (MCNN). The proposed model overcomes the limitations of shallow learning models and can determine a patient's survival rate and recommend personalized therapies. It processes a preprocessed dataset after missing value imputation and data normalization. Imputation fills in missing data, while normalization removes noise and outliers from the CKD datasetfrom the CKD dataset. The pre-processed data is passed through the convolution layer of the neural network for feature extraction, where matrix operations such as filters and strides create low-level feature maps. A max-pooling layer is applied to generate high-level feature maps, which are then passed to a fully connected layer for non-linear feature embedding. The embedded features are transformed into linear features through activation and batch normalization. These linear features are processed by a SoftMax function and a Naive Bayes classifier, effectively diagnosing and classifying the disease severity. A loss function is incorporated in the fully connected layer to minimize classification errors and improve accuracy. Experimental results show that the model achieves an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98.98{\text{\% }}$</tex>, outperforming state-of-the-art approaches, with performance evaluated through cross-fold validation and confusion matrix analysis.