Self-Attentive Semantic Distillation for Autonomous Criticality Mapping in Clinical Cyber Incident Reports

Abstract

The healthcare domain has increasingly become a target for cyber threats, with large-scale exposure of patient data and a steady rise in reported security vulnerabilities impacting hospital infrastructures. The rapid growth of unstructured medical security reports makes manual severity assessment inefficient, inconsistent, and unsuitable for timely response. To overcome these limitations, this study presents an automated text analysis framework based on Natural Language Processing (NLP), utilizing a medical security dataset composed of incident records, advisories, and vulnerability disclosures. The process begins with preprocessing and Exploratory Data Analysis (EDA) to ensure data quality through normalization, cleaning, and visualization. For contextual understanding, Lightweight RoBERT (Robustly Optimized BERT) is applied to generate semantic embeddings while maintaining computational efficiency. Unlike traditional approaches, the proposed system integrates Deep Neural Network (DNN)- based feature selection with Natural Gradient Boosting (NGBoost) to enhance classification performance. For comparison, baseline models such as Stochastic Gradient Descent (SGD) and NGBoost classifiers are also evaluated. The framework performs binary classification to distinguish between normal and highseverity vulnerabilities, enabling effective prioritization of critical issues. By combining contextual embeddings with advanced feature selection, the model improves accuracy, reduces false predictions, and enables faster response in real-time scenarios. This solution provides a scalable and efficient mechanism for automated vulnerability assessment, strengthening cybersecurity defenses and supporting improved risk management in healthcare systems