Deep Learning – Driven Sepsis Risk Forecasting with Composer –CNN – LSTM Integration

This study proposes a novel deep learning-based sepsis prediction framework, COMPOSER, which combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) units for more effective early and accurate detection of sepsis. This model was trained and validated on the MIMIC-III dataset, leveraging an optimized CNN-LSTM architecture, including a modified convolutional depth and a temporal modeling granularity, to maximize predictive accuracy. After tuning the number of CNN layers and temporal sequencing in the LSTM, the model reached the best classification metrics of a 91.1% accuracy, AUROC of 0.94, AUPRC of 0.89, precision of 87%, recall of 84%, and F1-score of 84%. These results show our approach to be a marked improvement to baseline models like MEWS, Logistic Regression, and standalone LSTM, especially in improving handling of class imbalance and false alarm count. The COMPOSER-enhanced CNN-LSTM variant did even better on interpretability and clinical applicability, while maintaining high deployment readiness and 4–6-hour early detection window, a factor crucial for proactive clinical intervention. The introduced framework addresses this concern by stacking several convolutional filters to have a richer perception of the local trends. LSTM gating mechanisms were deliberately crafted to better process long-term dependencies. The resulting COMPOSER architecture therefore represents a powerful, interpretable, high-performing clinical decision support system for early detection of sepsis in ICU environments.