M.S. AAI Capstone Chronicles 2024

Background Information Machine learning (ML) and deep learning (DL) models have emerged as powerful tools for the early detection and management of sepsis. These models leverage large datasets, often derived from electronic health records (EHRs), to identify patterns that may not be apparent to clinicians. By integrating patient demographics, vital signs, laboratory results, and other clinical parameters, ML and DL algorithms have shown some success in the ability to predict the onset of sepsis, enabling timely interventions that are critical for improving outcomes (Reyna et al., 2020). The COMPOSER model is an example of a real-time deep learning system that uses a feed-forward neural network to predict sepsis within four hours of onset (Boussina et al., 2024). It integrates multimodal data from EHRs, including vital signs, lab results, and patient demographics, and employs conformal prediction to minimize false positives, ensuring high-confidence alerts. TREWS, another widely used ML-based model, demonstrated reductions in mortality and organ failure through provider-validated sepsis alerts (Adams et al., 2022). Similarly, Sepsis Watch employs machine learning to provide early warnings but has yet to report direct patient-centered outcome improvements (Sendak et al., 2020). While these models show promise, they often face challenges such as high false positive rates, implementation complexity, and poor generalizability across diverse healthcare settings (Wong et al., 2021). Ongoing research in the adjacent field of brain-computer interfaces has yielded impressive results in interpreting multiple time-series signals from the brain and conducting classification tasks based on the interpretation of the composite of these signals (Zhao et al., 2024). These efforts have explored the efficacy of convolutional neural networks (CNN) and attention-based Transformer models to conduct this task. CNNs demonstrated success in

271