Exploration of deep learning-driven methods for monitoring abnormal vital signs in critical care patients
Abstract
The real-time recognition and interpretation of irregular physiological signals is a critical challenge in the development of intelligent healthcare systems, particularly within the domain of temporal health informatics where accurate analysis of time series data is essential. Traditional approaches often rely on static threshold-based methods, which fail to account for the temporal and inter-variable dependencies inherent in critical care data, thereby limiting their scalability and reliability in high-stakes clinical environments. To address these shortcomings, this study introduces a deep learning framework that integrates frequency-aware signal processing with context-driven attention mechanisms to model patient-specific abnormality patterns. Central to this framework is the ViSpecFormer architecture, a frequency-enhanced transformer model that leverages multiscale spectral filtering and temporal self-attention to capture both transient and persistent trends across multivariate signals. Additionally, the ClinConDec strategy dynamically adjusts abnormality thresholds based on clinical context, incorporating interpretable rule conditioning and episodic memory retrieval for transparent decision-making. Experimental evaluations on ICU datasets demonstrate that the proposed method outperforms conventional signal decomposition and autoregressive models in terms of accuracy and interpretability, particularly in scenarios characterized by heterogeneous clinical conditions and sparse observations. These findings highlight the potential of this approach to advance adaptive, reliable, and explainable physiological monitoring systems, contributing to the broader fields of human-centered computing, data-driven signal analysis, and interpretable machine learning.