Sequential Learning for Sepsis Prediction: LSTM and Transformer Architectures on ICU Time Series Data

Early detection of sepsis in ICU patients is critical because timely intervention can reduce mortality and improve clinical outcomes. This study presents a systematic framework for sepsis prediction using time-series data from ICU of over 40,000 patients. The main objective was to evaluate and compare sequential models, including LSTM, BiLSTM, and Transformer architectures, under the challenges of class imbalance and real-world ICU data, while carefully controlling for data leakage to ensure result validity. The models were trained with comprehensive preprocessing, including handling missing values, normalization, and noise reduction, and performance metrics such as precision, recall, F1-score, and AUROC were measured. The results showed that BiLSTM and Transformer architectures provide complementary strengths: BiLSTM achieves a balanced trade-off between precision and recall, whereas transformers better capture long-term temporal dependencies and early warning signals. This study demonstrates that the transition from classical machine learning to deep sequence models is not only a technical advancement but also a clinical necessity for reliable and early sepsis detection. All codes and preprocessing pipelines are publicly available to promote transparency and reproducibility.