Transformer-Based Framework for Predictive Patient Monitoring using Multi-Modal Data

Recently, predictive patient monitoring has played a crucial role in Intensive Care Units (ICUs), and the early detection of clinical weakening saves lives and removes hospital readmissions. However, the existing Artificial Neural Network with Particle Swarm Optimization (ANN-PSO) approach fails to capture complex temporal patterns in multi-modal data and struggles to integrate static patient details. This limits their ability to produce accurate and timely risk predictions of critical outcomes. To overcome these challenges, a transformer-based predictive framework for patient monitoring was proposed. Initially, data were collected from the MIMIC-III dataset and processed using cleaning to remove inconsistencies, normalization to scale numerical features, and segmentation to fixed-length time windows. These segmented data are then fed to the transformer model, which is designed to capture the complex temporal patterns. Here, each time step in a patient’s segment is transformed to embedding, and positional encoding is added to understand the order of events. These sequences are processed using a multihead self-attention mechanism, which enables the transformer to capture important temporal dependencies and feed-forward to produce a health profile. Furthermore, sigmoid activation was used for prediction, and the proposed transformer achieved better accuracy (99.21%), providing a robust framework for ICU patient monitoring.