Ventricular tachycardia (VT) is a life-threatening arrhythmia, but existing ICU monitors generate numerous false alarms, driven by motion artifacts, sensor disconnections, and electrical interference, which cause severe alarm fatigue. To address this, we study false VT alarm reduction using the VTaC benchmark, a dataset of over 5,000 expert-annotated ICU events comprising ECG, photoplethysmography (PLETH), and arterial blood pressure (ABP). We focus strictly on real-time detection, utilizing only the 10-second window immediately preceding the alarm onset.
Our approach utilizes a multi-modal 1D SE-ResNet classifier guided by two complementary inductive biases. First, we apply ICU-specific data augmentations (such as noise, baseline wander, and channel dropout) to simulate realistic monitoring failures. Second, we introduce auxiliary reconstruction tasks: PLETH is reconstructed via a data-driven decoder, while ABP utilizes a physics-informed differentiable forward simulation based on a three-element Windkessel model. By acting as a regularizer, this hemodynamic model penalizes latent representations that imply implausible pressure dynamics, discouraging the classifier from relying on non-physiological ECG artifacts.
This framework significantly outperforms prior real-time baselines, achieving a PhysioNet Challenge Score of 85.08 ± 1.65 on the VTaC test set (improving upon the previous best of 80.08 ± 2.46). Ablations reveal that while augmentations provide modest gains, physics-informed reconstruction consistently drives performance. We also identify corrupted PLETH as the most severe failure mode. Finally, Grad-CAM analysis demonstrates that our physics-informed model yields highly localized, clinically meaningful temporal activations. Ultimately, this work highlights that even a simple physiological forward model provides a powerful training signal for robust time-series classification.