Transformer vs. Manual Feature Engineering of Arterial Waveforms for Forecasting Hemodynamic Responses to Combined...

In the face of increased and more severe natural disasters around the world, future civilian and military medical evacuation (MEDEVAC) operations may require aggressive flight maneuvers, subjecting hemodynamically unstable casualties to potentially dangerous G-forces. Moreover, there are limited data and no known prediction models available to understand the physiological effects. Therefore, this paper presents the first set of models and a comparative analysis of transformer-based and traditional feature engineering approaches for forecasting hemodynamic responses under concurrent hemorrhage and G-force exposure. We developed a cross-dataset transfer learning framework using a modified Vision Transformer pre-trained on MIMIC-IV arterial blood pressure (ABP) waveform data and fine-tuned on experimental data that combines lower body negative pressure to simulate hemorrhage with centrifuge-induced G-force. Comparing zero-shot and fine-tuned transformer features against manually engineered ABP features, we found that while manual features achieved marginally better performance for forecasting systolic blood pressure (MSE: 30.19 ± 16.31 mmHg², R²: 0.84 ± 0.09), the fine-tuned transformer approach demonstrated comparable results (MSE: 32.43 ± 19.43 mmHg², R²: 0.82 ± 0.11) with no statistically significant difference. Principal component analysis revealed that manually engineered features captured more variance in fewer dimensions, while learned features showed promise in distinguishing between physiological conditions. This work establishes benchmarks for physiological forecasting under combined stressors and provides a foundation for developing autonomous MEDEVAC systems for operations in hostile environments.Clinical Relevance- Predicting hemodynamic responses under concurrent hemorrhage and G-force exposure may support the development of autonomous patient monitoring systems that improve real-time decision-making and casualty outcomes for medical air transport.