Online Sepsis Prediction Using Vital Signs and Multiscale Temporal-Aware Contrastive Learning: Model Development and Validation Study

Background Real-time prediction of sepsis is a critical yet highly challenging task. Existing studies face 2 major limitations. First, they often rely on laboratory test results that are not readily available in real time, making timely diagnosis difficult. Second, the patient’s condition evolves as a typical time series, but current methods often adopt coarse modeling strategies, with model architectures that are inefficient to train and deploy effectively. Objective This study aimed to develop a prediction model for online sepsis detection using only easily obtainable vital signs, such as heart rate and temperature, with variable-length input sequences while maintaining high predictive performance through the multiscale temporal representation learning. Methods We propose a deep learning model, Multi-Scale Temporal-aware Contrastive Learning (MSTCL), for efficient sepsis prediction based on the intensive care unit data derived from publicly available databases. We propose a multiscale temporal model to capture both short- and long-term dependencies in variable-length physiological time series. To enhance the robustness of our model, we used contrastive learning techniques that differentiate between positive and negative sepsis progression trajectories. Input features were limited to 6 vital signs, without reliance on laboratory tests or clinical notes. Results The model was evaluated on more than 400 patients with and without sepsis. It achieved an area under the receiver operating characteristic curve of 88.34%, a sensitivity of 89.29%, and a specificity of 73% for predicting sepsis onset based on variable-length vital-sign histories. The normalized mean absolute error for the predicted sepsis onset was 0.11%. Conclusions Our proposed model’s low complexity and rapid inference make it suitable for deployment in real-time monitoring systems and low-resource environments. The ability to learn from variable-length historical data enhances the clinical applicability of our model. Furthermore, the methodology of temporal-aware contrastive learning offers a robust and efficient solution for online sepsis detection in diverse clinical settings.