Leveraging High-Frequency Clinical Data for Early Septic Shock Detection with Advanced Machine Learning Frameworks

Septic shock, the most critical and life-threatening form of sepsis, poses a significant challenge in early detection, often resulting in high mortality, increased morbidity, substantial healthcare costs, and poor clinical outcomes. Timely intervention within critical hours is essential to improving prognosis, thereby highlighting the importance of accurate early prediction. In this study, we employed the HiRID dataset, a high time resolution ICU dataset that offers more frequent observations compared to the two widely used publicly available intensive care unit (ICU) datasets, MIMIC-III and eICU. To our knowledge, this is the first study to apply HiRID for septic shock prediction. Unlike previous studies that predominantly emphasize aggregated metrics, this study introduces a predictive model using machine learning to separately evaluate performance outcomes for both positive and negative classes, thereby increasing methodological transparency. In resource-limited critical care settings such as ICUs, frequent alarms can significantly impede clinical efficiency and timely patient care. Consequently, this study prioritizes the development of precise, short-term predictive models for five prediction windows (0.2, 0.4, 0.6, 0.8, and 1.0 hours). These short intervals were deliberately selected to facilitate immediate and actionable interventions, rather than forecasting several hours in advance, which may offer less clinical utility. We trained ten machine learning models to forecast septic shock using stratified 5-fold cross-validation with strict patient-level data separation to ensure robustness and prevent data leakage. For the 0.2-hour onset prediction task, our best-performing fold achieved perfect scores across all key metrics—accuracy, AUC, precision, recall, and F1-score. Even in the least favourable fold, the model demonstrated robust performance with 100% accuracy, an AUC of 0.97, class-wise precision of (1.00, 1.00), recall of (1.00, 0.91), and F1-scores of (1.00, 0.95), indicating consistent and reliable predictive capability. These results underscore the effectiveness and clinical utility of our framework for early septic shock prediction and its potential to support timely, actionable decision-making in critical care settings.