EARLY SEPSIS DETECTION FROM ICU DATA USING WAVELET FEATURE EXTRACTION WITH HYBRID DEEP LEARNING FRAMEWORK

For patients in intensive care units (ICUs) to have better outcomes after suffering from sepsis, a potentially fatal illness, prompt and precise identification is essential. Timely care and better patient outcomes depend on early sepsis detection. It is common for traditional diagnostic procedures to miss the early signs of sepsis, which means that treatment is delayed. Using the open-source MIMIC-III clinical database, we provide a new hybrid diagnostic system for early sepsis identification that makes use of deep learning, signal processing, and metaheuristic optimisation. We start by using physiological time-series data including temperature, respiration rate, blood pressure, and heart rate to extract robust, multi-scale, and noise-resistant features using the Wavelet Scattering Transform (WST). Afterwards, a one-dimensional input-adapted AlexNet-based Convolutional Neural Network is trained with these features to learn the high-level feature representations necessary to differentiate among septic and non-septic patients. We optimise hyper parameters including learning rate, filter sizes, and dropout rates using the Hunger Games Search (HGS) method, a new metaheuristic inspired by nature, to further enhance model accuracy and generalisation. Benchmark CNN architectures and traditional machine learning models are outperformed by the suggested WST-AlexNet-HGS pipeline. Classification performance above 95% as shown by evaluation measures like accuracy, precision, recall, and F1-score, demonstrating strong diagnostic reliability. Given the model's efficacy, it might be used as a clinical decision support tool in intensive care units operating in real-time. This hybrid pipeline demonstrates >95% accuracy in early detection of sepsis from ICU data, outperforming conventional methods. The model's robustness and reliability position it as a viable decision support system for real-time clinical use.