Intelligent Sepsis Detection in ICU: A Real-Time Deep Learning Framework with Hybrid Optimization

Without rapid medical intervention, sepsis—the body's catastrophic reaction to infection—can cause organ failure and increased mortality. There is an immediate need for real-time and automated detection methods because conventional diagnostic methods are often inaccurate or have delays. In this research, to present a deep learning-based sepsis detection system that uses a CNN for early prediction in conjunction with a Modified Bidirectional Gated Recurrent Unit (MBiGRU). Convolutional neural networks (CNNs) extract spatially rich and clinically significant features, while MBiGRUs efficiently model temporal patterns in physiological time-series data. The framework incorporates a modified Chaotic Zebra Optimization Algorithm to optimise feature selection and generalization, significantly enhancing model efficiency and eliminating redundant information. Using open-source sepsis datasets, to trained and validated the MBiGRU-CNN-CZOA model. In comparison to more conventional models such simple CNN, LSTM, and GRU, the model outperformed them with a 93.82% accuracy, 94.24% precision, 93.82% recall, and 93.68% F1-score. Fast decision-making and efficient calculation are guaranteed by the system's optimisation for real-time ICU deployment through edge computing. Based on these results, the suggested approach seems to be well-suited for clinical use in the real world, where it could facilitate prompt intervention and drastically cut down on mortality risks. In order to ensure the security of patient data when used in remote healthcare settings, future research will centre on improving interpretability and implementing federated learning strategies.