Abstract Body (Do not enter title and authors here): Introduction: Ventricular tachycardia (VT) is a life-threatening clinical condition and common cause of cardiac arrest in the intensive care unit (ICU) setting. The current ICU-VT clinical care model is reactionary, determined by alarm-triggered notifications following VT onset involving poorly performing R-peak-dependent features subject to waveform distortion. Timeseries foundation models (TFMs) circumvent these limitations by learning intrinsic patterns directly from raw electrocardiogram (ECG) data. We hypothesized that TFM-based VT event detection reduces false alarms while maintaining high sensitivity for true VT events.

Methods: We analyzed previously adjudicated VT alarms events from PhysioNet VTaC dataset. A board-certified electrophysiologist (T.D.) further annotated exact VT onset and offset event times from N=1241 cases. ECG lead II waveforms were segmented into 512-data-point windows and labeled as VT-positive if >75% of the segment overlapped with annotated VT. These segments were processed using MOMENT, a TFM pretrained on diverse timeseries data to generate representations. The resulting features were used for training a Support Vector Machine classifier for VT. For testing, we evaluated N=206 bedside monitor-reported VT alarms (06/2024-11/2024) from the University of Maryland Medical Center, adjudicated by T.D., resulting in N=31 (15%) true and N=175 false (85%) VT events.

Results: In the testing dataset, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.79 (95% confidence interval 0.75-0.83). At the optimal operating threshold, sensitivity was 74.2% (N=23/31 true VT detected) with 84.6%, specificity (N=148/175 non-VT events identified correctly). Negative predictive value (NPV) was 95.1% vs. positive predictive value (PPV) of 46.1%. The absolute false alarm rate decreased from 85.0% to 15.3%.

Conclusions: We developed a TFM that enables accurate VT detection with generalizable waveform representations, and rules out false alarms effectively toward reducing alarm fatigue while ensuring timely response to true VT events. Future work integrating data from multiple patient-level waveforms (e.g. blood pressure, others) and 1 ECG lead is expected to optimize true VT event detection further and establish a refined TFM that is clinically actionable for ICU patients.