Embedding-Enhanced Multimodal Siamese Network for False Arrhythmia Alarm Reduction in Intensive Care Units

False arrhythmia alarms in ICUs remain a major issue, with up to 88.8 % being false. These alarms are often triggered by electrode noise due to patient movement, power line interference, contact artifacts, device-related distortions, or signal-clinical discordance. This alarm burden contributes to staff fatigue, reduced responsiveness, and increased patient stress. Although intelligent systems, driven by recent advances in machine learning and deep learning, are increasingly used to support clinical decisions, they have yet to significantly improve alarm validation. With only 2 % to 9 % of alarms requiring intervention, this underscores the need for more effective models to distinguish critical events from false alarms. This work proposes a deep learning approach that combines a multimodal architecture with a Siamese network to detect false arrhythmia alarms. The model is trained using raw ECG, ABP, and PPG signals from the PhysioNet/CinC Challenge 2015 dataset. The multimodal design enables joint signal processing, while the Siamese structure improves similarity-based representation learning. The proposed method achieves a Challenge score of 71.14%, surpassing the most comparable prior method (60.69%) by 10.45%. These re-sults demonstrate the potential of embedding-enhanced Siamese networks for improving alarm accuracy in critical care.