Physics-Informed Deep Learning for False Ventricular Tachycardia Alarm Reduction in the ICU
The paper says a physics-guided auxiliary task cut false VT alarms better than the prior VTaC benchmark leader.
The model pairs a 1D SE-ResNet with ICU-style augmentations and a differentiable Windkessel simulation that reconstructs plausible arterial pressure waveforms. The authors argue this penalizes ECG artifacts while preserving true ventricular tachycardia signals across modalities. Under a 10-second pre-alarm real-time protocol, it reports a 5-point Challenge Score gain over prior state of the art. Ablations identify the physics-informed objective as the main driver, with better accuracy, twice the label efficiency, and more clinically localized ECG evidence. ArXiv · AI/CL/LG's note
The model pairs a 1D SE-ResNet with ICU-style augmentations and a differentiable Windkessel simulation that reconstructs plausible arterial pressure waveforms. The authors argue this penalizes ECG artifacts while preserving true ventricular tachycardia signals across modalities. Under a 10-second pre-alarm real-time protocol, it reports a 5-point Challenge Score gain over prior state of the art. Ablations identify the physics-informed objective as the main driver, with better accuracy, twice the label efficiency, and more clinically localized ECG evidence. ArXiv · AI/CL/LG's note
score 4