Fine-tuned U-Net and time-weighted bi-directional LSTM for low-latency cardiopulmonary monitoring in WBSN data

The evolution of Wireless Body Sensor Networks (WBSNs) has dramatically changed the Smart Emergency Monitoring (SEM) landscape, enabling continuous and non-invasive monitoring of vital signs such as blood pressure, SpO 2 , respiratory rate, and even CO 2 levels. However, the timely and precise detection of transient yet clinically significant events remain an enormous statistical challenge owing to the lower sensitivity and specificity offered by traditional deep learning methods. Such salient experimental alerts usually go with a higher error in prediction rate, hence generalizing poorly with sub-average temporal context modeling from U-Net architectures combined with LSTM layers. To defeat such weaknesses, it proposed a statistically refined hybrid approach, followed by the fine-tuned U-Net framework for spatial feature extraction, interfaced with a Time-Weighted Multi-Task Bi-Directional LSTM (TW-MT-BiLSTM) for temporal sequence modeling. Three enhancements are considered statistically significant, such as: (1) attention-based temporal weighting limiting higher consideration to important time segments, (2) bidirectional LSTM giving better inference capitalizing on past and future contexts, and (3) multi-task learning simultaneously predicting the variables of blood pressure, SpO 2 , and respiratory rate, which decreases variance and enhances the robustness of overall prediction. An extensive set of experiments on benchmark datasets such as Medical Information Mart for Intensive Care (MIMIC-III) and PhysioNet is compared in terms of statistical performance metrics, including Mean Absolute Error (MAE), Root Mean Square, and the superiority of the proposed methods was revealed over existing methods.