Sepsis is a life-threatening infection affecting 50 million people annually, with survival decreasing by up to 8% per hour of delayed treatment. Current diagnostics rely on blood culture, which takes days and delays targeted therapy. Our recent work demonstrated a culture-free method combining smart centrifugation, microfluidic bacterial trapping, and time-lapse microscopy, enabling rapid bacterial detection directly from blood (https://www.biorxiv.org/content/10.1101/2024.05.23.595289v2.abstract). Deep learning models successfully identified the presence of bacteria within two hours at clinically relevant concentrations (1–10 CFU/ml). This project aims to extend detection to species-level identification using deep learning. We will: 1. Collect and preprocess large time-lapse microscopy datasets of multiple clinically relevant species (E. coli, K. pneumoniae, E. faecalis, S. aureus). 2. Develop and train CNN and transformer-based models (e.g., EfficientNet, DinoV2, Video ResNet) for multi-species classification from microfluidic trap images and time-lapse sequences. 3. Evaluate both single-frame and video-based classification to enable robust, real-time inference. GPU computing is essential due to the high-dimensional imaging data, tens of thousands of time-lapse sequences, and the computational demands of modern deep learning models. Resources on Alvis (C3SE) with Mimer storage will support large-scale training, hyperparameter optimization, and deployment. The expected outcome is a rapid, species-level diagnostic tool that eliminates the need for blood culture, reducing diagnostic time from days to hours. This will enable earlier targeted therapy, limit broad-spectrum antibiotic use, and improve sepsis patient outcomes while advancing AI-driven clinical microscopy.