Physics-Informed Deep Learning for False Ventricular Tachycardia Alarm Reduction in the ICU

1. Introduction

• Sensor detachment, patient movement, and electrical interference can produce ECG artifacts that closely resemble true arrhythmias, confounding data-driven classifiers.VTaC provides over 5,000 multi-institutional ICU recordings with ECG, PLETH, and ABP for evaluation.Source: §1 · p. 1
• Prior real-time methods achieved Challenge Scores up to 80.08, while higher-AUC foundation-model and retrospective approaches rely on external pretraining or post-alarm information.These approaches remain vulnerable to severe sensor noise and previously unseen artifact patterns.Source: §1 · p. 1
• 85.08 ± 1.65 Challenge Score, a ∼5-point improvement over prior state-of-the-art, using a strict 10 s real-time window without external pretraining data.The method jointly optimizes classification with physics-informed ABP reconstruction and data-driven PLETH reconstruction.Source: §1 · p. 1

2.1. Dataset and Preprocessing

• 5,037 expert-annotated VTaC alarm events from three U.S. hospitals are split patient-wise 80-10-10, using only the final 10 s before each alarm.Fourteen canonical channels are mapped to fixed slots with zero-filled unavailable channels and a concatenated availability mask; each channel is independently z-normalized.Source: §2.1 · p. 1

2.2. Architecture

• The 1D SE-ResNet1D backbone uses a three-branch multi-scale stem and four feature stages, with shared features feeding classification and ABP and PLETH reconstruction heads.Models contain 2–9M trainable parameters depending on auxiliary heads; hyperparameters and the decision threshold are selected using validation data.Source: §2.2 · pp. 1–2, Figure 1 · p. 2

2.3. Augmentation and Class Imbalance

• Five ICU-oriented augmentations target signal degradation, while Asymmetric Focal Loss addresses class imbalance using a tunable false-negative weight.Augmentations include temporal jitter, Gaussian noise, channel dropout, amplitude scaling, and 0.1–0.5 Hz baseline wander.Source: §2.3 · p. 2

2.4. Physics-Informed Regularization

• A six-parameter projection head drives a differentiable two-stage Windkessel simulation to reconstruct ABP as an auxiliary task, not as the model’s prediction target.A synthetic ejection proxy feeds the three-element model, whose exact exponential integration is stable for positive Rp and C.Source: §2.4 · p. 2
• The ABP reconstruction loss regularizes shared features because artifact-driven representations are less compatible with coherent hemodynamics and incur higher reconstruction error.PLETH is reconstructed by a data-driven MLP; its availability is about 91% of recordings versus about 36% for ABP, and reconstruction losses are masked when channels are unavailable.Source: §2.4 · pp. 2–3

3. Results

• The primary Challenge Score is (TP + TN)/(TP + TN + FP + 5 · FN), weighting missed true alarms fivefold; reported scores are mean ± SD over five random seeds.Source: §3 · p. 3

3.1. Main Results

• Table 1 compares model variants with prior work on VTaC using Challenge Score, reported as mean ± SD over five seeds where available.The table caption identifies augmentation, data-driven reconstruction, and physics-informed reconstruction variants.Source: Table 1 · p. 3, §3.1 · p. 3

3.2. Ablation Studies

• Without augmentations, physics-informed reconstruction improves the baseline by +1.42 Challenge Score points versus +0.57 for data-driven reconstruction; with augmentations, it improves over augmentation alone by +3.37 versus +2.45.Source: §3.2 · p. 3
• At 5% labels, the physics-informed model scores 64.74±2.94, nearly matching the augmentation-only baseline at 10% labels (65.04 ± 1.98), and is the only multi-task variant to beat baseline at every fraction.Table 2 evaluates label fractions of 5%, 10%, 20%, and 50%.Source: §3.2 · p. 3, Table 2 · p. 3
• Sensor dropout reduces Challenge Score by −1.33 for ABP and −4.85 for PLETH; replacing signals with uncorrelated Gaussian noise reduces it by −4.98 for ABP and −6.42 for PLETH.Corrupted signals are reported as more damaging than missing signals; Table 2’s modality-robustness results are from a single run.Source: §3.2 · p. 3, Table 2 · p. 3
• On 128 true-alarm samples, the physics-informed model produces more localized Grad-CAM saliency in 89.1% of cases, with mean Gini coefficient rising from 0.52 ± 0.11 to 0.65 ± 0.09.Figure 2 illustrates a false alarm that the baseline misclassifies but the physics-informed model correctly rejects.Source: §3.2 · p. 3, Figure 2 · p. 3

• Sensor detachment, movement, and electrical interference can create wide-complex, high-rate ECG patterns indistinguishable from VT on a single lead.Source: §4 · p. 3
• The Windkessel head penalizes artifact-driven features because electrical artifacts lack a hemodynamic correlate, making simulated pressure incoherent and increasing reconstruction loss.Source: §4 · p. 4
• ABP reconstruction loss decreases only approximately 2—3%, and the Windkessel model coarsely approximates measured waveforms; limitations also include ABP in 36% of recordings and evaluation on one benchmark.Despite modest reconstruction accuracy, the fixed dynamical structure propagates corrective gradients, unlike a flexible MLP decoder that could absorb artifacts.Source: §4 · p. 4

• Embedding a simple hemodynamic forward model as an auxiliary reconstruction task provides a powerful inductive bias for false VT alarm reduction.Source: §5 · p. 4