Real-Time Multivariate Vital Sign Forecasting in Intensive Care Units: A Comparative Study of Machine Learning Models with Emphasis on Time Series Mixer for Early Warning Systems

Real-time monitoring and prediction of physiological parameters in critical care settings remains essential for preventing patient deterioration and enabling timely medical interventions. This research examines various computational approaches for short-term prediction of six critical vital signs: Heart Rate (HR), Systolic/Diastolic Blood Pressure (SBP/DBP), Respiratory Rate (RR), Oxygen Saturation (SpO2), and Temperature (Temp) using the VitalDB clinical database. Our investigation began with conventional time-series methods including Autoregressive Integrated Moving Average (ARIMA) models and progressed to sophisticated neural architectures such as Long Short-Term Memory (LSTM) networks. However, these approaches demonstrated limitations in modeling complex relationships between multiple physiological variables simultaneously. Subsequently, we implemented advanced hybrid architectures incorporating Bidirectional LSTM (BiLSTM) layers, Convolutional Neural Networks (CNN), and Graph Attention Network (GAT) mechanisms. Although this hybrid model achieved enhanced prediction accuracy, their computational complexity posed challenges for clinical deployment. Addressing practical implementation requirements, we evaluated Multi-Layer Perceptron (MLP)-based frameworks, specifically Patch Time Series Transformer (PatchTST) and Time Series Mixer (TSMixer) architectures. PatchTST effectively captures extended temporal dependencies but lacks comprehensive cross- variable interaction modeling. Conversely, TSMixer employs dual mixing mechanisms—temporal and feature-based—to simultaneously learn chronological patterns and inter-vital correlations. Utilizing 10-minute historical windows to forecast subsequent 3-minute intervals, TSMixer demonstrated superior performance across all evaluation metrics. The model achieved the lowest Root Mean Square Error (RMSE) values for all vital parameters while maintaining computational efficiency suitable for real-time applications. These findings establish TSMixer's potential as a practical solution for prospective integration into Intensive Care Unit (ICU) monitoring systems, offering both predictive accuracy and operational feasibility for clinical environments.