A Hybrid Deep-Boosting Framework for Context-Aware Severity Classification in Healthcare Security Narratives

Abstract

Healthcare infrastructures are becoming prime targets for cyber threats, as evidenced by the massive exposure of patient records and the continuous rise in reported vulnerabilities within hospital systems. The growing dependence on digital platforms has intensified the need for efficient mechanisms to detect and prioritize security issues. Traditional manual methods for assessing bug severity are not only time-intensive but also inconsistent, particularly when dealing with large volumes of unstructured security reports generated in medical environments. To address these challenges, this research introduces an intelligent Natural Language Processing (NLP)-based framework designed to automate bug severity classification. The approach utilizes a specialized dataset containing medical incident reports, security advisories, and vulnerability disclosures. Initially, the data undergoes systematic preprocessing and Exploratory Data Analysis to ensure cleanliness, consistency, and meaningful representation. For capturing contextual semantics, a lightweight transformer-based model inspired by Robustly Optimized bidirectional encoded representation of transformers (RoBERTa) is employed, enabling efficient generation of text embeddings with reduced computational overhead. To further enhance predictive performance, the framework integrates a Deep Neural Network (DNN) for feature refinement, which identifies and extracts the most relevant patterns from the embedding space. These optimized features are then passed to a probabilistic boostingbased classifier, improving classification reliability compared to conventional standalone models. The system categorizes vulnerabilities into two classes normal and severe facilitating rapid prioritization of critical issues. The framework demonstrates improved accuracy, reduced misclassification, and faster processing capability. By combining efficient language modelling with advanced feature learning, this approach offers a scalable and practical solution for strengthening cybersecurity management in healthcare environments.