Abstract
Sepsis is a life threatening condition that requires timely detection inintensive care settings. Traditional machine learning approaches, includingNaive Bayes, Support Vector Machine (SVM), Random Forest, and XGBoost, oftenrely on manual feature engineering and struggle with irregular, incompletetime-series data commonly present in electronic health records. We introduce anend-to-end deep learning framework integrating an unsupervised autoencoder forautomatic feature extraction with a multilayer perceptron classifier for binarysepsis risk prediction. To enhance clinical applicability, we implement acustomized down sampling strategy that extracts high information densitysegments during training and a non-overlapping dynamic sliding window mechanismfor real-time inference. Preprocessed time series data are represented as fixeddimension vectors with explicit missingness indicators, mitigating bias andnoise. We validate our approach on three ICU cohorts. Our end-to-end modelachieves accuracies of 74.6 percent, 80.6 percent, and 93.5 percent,respectively, consistently outperforming traditional machine learningbaselines. These results demonstrate the framework's superior robustness,generalizability, and clinical utility for early sepsis detection acrossheterogeneous ICU environments.