Sepsight : Sepsis prediction from Medical Data using Transformer Model

Sepsis is a life-threatening condition requiring early detection for improved patient outcomes. Traditional prediction methods, including SIRS, qSOFA, and LSTMs, struggle with capturing complex temporal dependencies, limiting real-time prediction accuracy. This research introduces a decoder-only transformer model for sepsis prediction using EMR time series data. The self-attention mechanism effectively identifies critical moments in patient data, capturing intricate temporal patterns and long-range dependencies. The architecture integrates data tokenization, positional encoding, and masked multi-head self-attention for dynamic analysis. Dropout regularization and hyperparameter tuning enhance robustness and prevent overfitting. Experimental results demonstrate superior accuracy, sensitivity, and specificity compared to traditional approaches. The model’s parallel processing capability enables scalability, allowing real-time assessments across large datasets. Additionally, attention mechanisms improve interpretability, aiding clinical decision- making. Beyond sepsis detection, this research highlights the potential of transformer models in predictive healthcare. Their ability to process sequential medical data efficiently can enhance disease monitoring, risk assessment, and early intervention strategies.