Hemorrhagic shock remains a leading cause of preventable death in trauma patients, with delayed recognition of early blood loss potentially contributing to increased mortality. Traditional monitoring methods based on vital signs and laboratory tests f...
Hemorrhagic shock remains a leading cause of preventable death in trauma patients, with delayed recognition of early blood loss potentially contributing to increased mortality. Traditional monitoring methods based on vital signs and laboratory tests frequently lack sensitivity in detecting and assessing the severity of hemorrhage. These limitations become particularly pronounced—and are poorly understood—under conditions of physiological compensation or pharmacological suppression, such as those induced by β-blocker administration. To overcome these limitations and enable earlier detection compared to conventional methods, we developed a hybrid deep learning model combining one-dimensional convolutional neural network (1D CNN) and bidirectional long short-term memory (BiLSTM) layers for real-time estimation of hemorrhage progression and compensatory reserve from multichannel physiological signals collected in a controlled swine hemorrhage-resuscitation model, including subjects receiving the β-blocker.
Eight physiological signals—heart rate (HR), mean arterial pressure (MAP), end-tidal carbon dioxide (EtCO2), arterial blood pressure (ABP), photoplethysmogram (PPG), electrocardiogram (ECG), pulse pressure variation (PPV), and central venous pressure (CVP)—were continuously recorded and analyzed using multitask modeling to detect bleeding events, classify hemorrhage severity, identify β-blockade status, estimate cumulative blood loss volume (BLV), and predict compensatory reserve index (CRI). The hemorrhage detection model achieved an AUC of 0.96, reliably identifying bleeding states even under β-blockade conditions. Hemorrhage class classification showed an overall accuracy of 72.3%, which is likely due to the subtle physiological differences between early hemorrhage phases and their temporal proximity between hemorrhage classes. The model effectively identified β-blocker treated subjects during both hemorrhage and resuscitation phases (AUC = 0.99), confirming its sensitivity to pharmacologically altered responses.
For regression tasks, continuous BLV estimation achieved strong correlation (r = 0.74), although slightly overestimated blood loss under β-blockade, reflecting clinically meaningful changes in compensatory dynamics relevant for early intervention decisions. Continuous CRI prediction also demonstrated robust predictive performance (r = 0.81), closely approximating the dynamic interplay of physiological compensation and decompensation.
These findings demonstrate that the proposed deep learning framework effectively integrates multiple physiological signals, capturing subtle, clinically relevant physiological variations and enabling real-time, clinically actionable insights. This approach shows strong potential for early hemorrhage detection and management, especially in emergency or resource-limited settings, and maintains robust performance even under physiologically altered conditions such as β-blocker administration. Thus, the proposed framework provides a quantitative foundation for capturing individualized physiological responses, and is expected to contribute meaningfully to the development of patient-specific therapeutic strategies in future clinical applications.