Abstract
Sepsis is a life-threatening syndrome characterized by severe organ dysfunction in response to infection. It remains a leading cause of mortality in Intensive Care Units (ICUs) worldwide. While early administration of antibiotics can significantly reduce complications, anticipating sepsis onset remains a major clinical challenge. Current machine learning models often suffer from unrealistic evaluation frameworks—such as requiring a priori knowledge of onset times—or fail to handle the extreme class imbalance inherent in daily clinical screening. In this paper, we propose a novel prediction framework that utilizes data collected during the night (10 PM – 6 AM) to predict sepsis onset within the following 24 hours. To address the resulting class imbalance, we introduce Nightly Profile Representation Learning (NPRL), a self-supervised approach that preserves the diversity of patient data. Our empirical study using data from a level-1 trauma center demonstrates that NPRL significantly improves early detection performance over standard baselines.
Introduction
In the ICU, every hour matters. Sepsis is a medical emergency where the body’s immune response to infection causes systemic damage. For trauma patients, the diagnostic challenge is even greater: the physiological stress of injury often mimics the early signs of sepsis, such as inflammation and elevated heart rate, making it difficult for clinicians to distinguish between recovery and a developing infection.
While deep learning has shown promise in medical diagnostics, existing sepsis prediction models often fall into two traps. First, many use a "fixed time to onset" training window, which assumes the time of sepsis is already known—a luxury not available in real-time care. Second, when models are deployed for daily screening, they encounter a "rare event problem." Because sepsis occurs in a small fraction of patient-days, standard models often default to predicting "no sepsis" for everyone to achieve high nominal accuracy, rendering them useless for clinical intervention.
The NPRL Framework
To bridge the gap between research and clinical reality, we developed the Nightly Profile Representation Learning (NPRL) framework. This approach is built on two key pillars: a realistic prediction window and a self-supervised learning strategy to combat data imbalance.
Strategic Timing: The Nightly Window
Our model mimics a daily diagnostic test. It analyzes patient data from the "quiet" night shift (10 PM to 6 AM) to predict whether sepsis will occur within the next 24 hours. This timing is clinically strategic for two reasons:
- Staffing Ratios: Night shifts typically have higher patient-to-provider ratios. A model that processes data overnight and provides a risk score by 6 AM can act as a "pre-screening assistant" for clinicians during morning rounds.
- Signal Clarity: The nighttime period often features fewer external interruptions, diagnostic procedures, and interventions, potentially offering a "cleaner" physiological signal of the patient's underlying state.
Overcoming Imbalance via Representation Learning
The core innovation of NPRL is its use of self-supervised pre-training. Before the model ever sees a "sepsis" label, it is trained to uniquely identify and distinguish between individual "patient-night" profiles. Mathematically, we prove that by forcing the model to learn a diverse latent space that captures the unique characteristics of every patient-night, we prevent the model's weights from collapsing toward the majority (non-sepsis) class. When the model is later fine-tuned for sepsis prediction, it retains this sensitivity to rare, subtle physiological deviations.
Methodology
We utilized a Multi-modal RNN architecture to process the complex data streams found in the EHR:
- Temporal Data: A Bidirectional Gated Recurrent Unit (GRU) processes time-series data, including hourly vitals and laboratory results.
- Static Data: A deep dense network handles non-temporal information, such as age, injury type, and initial Emergency Department physiology.
- Cumulative Exposure: The model also accounts for cumulative factors like total IV fluid volume and days spent on a ventilator, which our analysis found to be highly predictive of onset.
Experimental Results
The study was conducted using de-identified EHR data from UW Medicine Harborview Medical Center, involving 2,802 trauma patients. While 17% of these patients eventually developed sepsis, the daily instance rate was less than 2%, representing a severe class imbalance.
NPRL demonstrated superior performance compared to traditional machine learning (XGBoost) and standard recurrent neural networks (RNNs):
- AUROC: NPRL achieved 0.7870, outperforming the standard RNN (0.7794) and XGBoost (0.7398).
- Sensitivity (Recall): Most importantly for a screening tool, NPRL reached a sensitivity of 0.8280, catching nearly 83% of sepsis cases. This was a significant improvement over the standard RNN's sensitivity of 0.7346.
Discussion and Clinical Utility
By delivering high-sensitivity predictions at 6 AM, NPRL provides actionable intelligence when staffing is most prepared to initiate diagnostic workups or adjust treatment plans.
However, the model does face a trade-off in specificity (0.6120). While it is excellent at catching sepsis, it generates more false positives than some conservative baselines. In an ICU environment, this must be balanced against "alarm fatigue." Nevertheless, in the context of a life-threatening condition like sepsis, we argue that high sensitivity is the primary requirement for an effective early-warning system.
Conclusion
NPRL offers a more realistic and robust approach to sepsis prediction by focusing on the nightly window and utilizing self-supervised representation learning to overcome the rare event problem. Future work will involve validating this model on multi-center datasets and exploring data augmentation techniques to further refine the representation of patient profiles in electronic health records.
Summary updated