Revolutionizing Healthcare: Comprehensive Evaluation and Optimization of SVM Kernels for Precise General Health Diagnosis

Abstract

Purpose: This study is driven by a two-fold objective. Firstly, it seeks to optimize the Support Vector Machine (SVM) algorithms in machine learning, comprehensively evaluating diverse SVM kernel variants to enhance their versatility and applicability across various domains, which is beyond the healthcare sector. Secondly, in the context of general health diagnosis, it aims to assess the suitability of SVM kernels for achieving precision in predictive modeling. The choice of SVM is rooted in its effectiveness, proven in classification and regression within data mining. SVMs excel in high-dimensional problem classification, demonstrating superior accuracy, making them invaluable in refining machine learning methodologies and advancing diagnostic systems, promising implications for healthcare and beyond. The chosen SVM model, distinguished by its exceptional performance, is then implemented in real-world applications, particularly in wireless, non-invasive healthcare devices. This deployment signifies a substantial stride toward advancing healthcare practices and holds promising implications for various fields.

Methods: Data for this study was collected from publicly accessible datasets on Kaggle, encompassing a comprehensive array of general health-related information. This dataset, comprised of clinical data and vital signs data, underwent meticulous preprocessing, such as data cleaning, feature extraction, and categorization of health status into ‘healthy’ and ‘requiring further attention’. Subsequently, predictive models were constructed employing Support Vector Machine (SVM) algorithms with various kernel functions, such as Linear, RBF, Polynomial, and Sigmoid. They were trained and tested on the preprocessed dataset to assess their efficacy in general health diagnosis. Model performance was rigorously evaluated using established metrics, including accuracy, precision, recall, F-1 score, Area Under the Curve (AUC), Receiver Operating Characteristic (ROC) curve, and cross-validation. The selection of the most efficacious SVM kernel was governed by stringent adherence to industry standards and best practices, ensuring optimal integration into health diagnostic systems. The chosen model was tested using new datasets obtained from wireless non-invasive healthcare devices and the pre-existing AHD application. Hyperparameter tuning was meticulously executed to maximize accuracy, ensuring the effectiveness of the evaluation process.

Results: The results demonstrate that the Polynomial kernel was selected as the body health diagnostic model instead of the Linear, RBF, and Sigmoid kernels. This kernel has a training time of 0.8 seconds, a testing time of 0.1 seconds, accuracy scores of 97%, precision of 97%, recall of 97%, F-1 score of 97% for training metrics, and accuracy scores of 99%, precision of 99%, recall of 99%, and F-1 score of 99% for testing metrics. The accuracy of the polynomial kernel model decreased to 0.88 on new datasets; adjusting the hyperparameter C to C = 100 resulted in the highest accuracy of 0.945.

Novelty: This study introduces a pioneering approach by rigorously optimizing Support Vector Machine (SVM) algorithms, notably the innovative application of the Polynomial kernel in general health diagnosis. Unlike traditional kernels, the Polynomial kernel exhibited exceptional accuracy (up to 99%) and precision. Furthermore, the study’s unique methodology, combining industry standards and meticulous hyperparameter tuning, ensures seamless integration into real-world healthcare systems. The deployment of this optimized model in wireless non-invasive healthcare devices signifies a groundbreaking advancement, highlighting a novel synthesis of theoretical innovation and practical implementation in machine learning for healthcare.