Andrew Heroy, MSc
Uniformed Services University of Health Sciences/ Metis Foundation
"Hemorrhagic shock (HS) remains a preventable cause of death, with casualties deteriorating before surgical control despite point-of-injury care. In a porcine model (n=12) of controlled hemorrhage, we classified HS by percentage blood volume loss: C1 (85-100%), C2 (70-85%), C3 (60-70%), C4 (< 60%).
Methods
A controlled bleed rate was used to define the four HS stages. We developed a Signal Quality Index (SQI) using Welch’s method to estimate Power Spectral Density (PSD). Two metrics gated the signal: (1) In-Band Power Ratio (0.5–15 Hz) and (2) Normalized Shannon Entropy. Segments failing thresholds (in-band power 0.50) were excluded. 43 features were extracted. Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), XGBoost and Voting Classifier (VC) models were evaluated via Leave-One-Subject-Out cross-validation. A mechanistic Windkessel ODE model is also being developed to provide physiological context for carotid flow.
Results
The models achieved the following mean classification accuracies: SVM (43.80% +/- 17.14%), KNN (45.15% +/- 12.84%), VC (54.43% +/- 13.30%), RF (54.95% +/- 15.06%), and XGBoost (58.02% +/- 16.91%). The parallel ODE approach is in early development but is vital for prediction improvement by mapping the underlying physiology.
Conclusions
Automated noise rejection is essential for robust classification of HS from hemodynamic waveforms. Our research offers an explainable path toward real-time triage."
Methods
A controlled bleed rate was used to define the four HS stages. We developed a Signal Quality Index (SQI) using Welch’s method to estimate Power Spectral Density (PSD). Two metrics gated the signal: (1) In-Band Power Ratio (0.5–15 Hz) and (2) Normalized Shannon Entropy. Segments failing thresholds (in-band power 0.50) were excluded. 43 features were extracted. Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), XGBoost and Voting Classifier (VC) models were evaluated via Leave-One-Subject-Out cross-validation. A mechanistic Windkessel ODE model is also being developed to provide physiological context for carotid flow.
Results
The models achieved the following mean classification accuracies: SVM (43.80% +/- 17.14%), KNN (45.15% +/- 12.84%), VC (54.43% +/- 13.30%), RF (54.95% +/- 15.06%), and XGBoost (58.02% +/- 16.91%). The parallel ODE approach is in early development but is vital for prediction improvement by mapping the underlying physiology.
Conclusions
Automated noise rejection is essential for robust classification of HS from hemodynamic waveforms. Our research offers an explainable path toward real-time triage."
