Project Summary:
Sepsis is a life-threatening condition characterized by a dysregulated immune response to infection, leading to
systemic inflammation and multi-organ dysfunction. Despite advances in medical care, sepsis remains a
significant cause of morbidity and mortality worldwide, with early detection and intervention being critical to
improving patient outcomes. However, the early detection of sepsis poses several challenges due to its
heterogeneous presentation, non-specific symptoms, and the dynamic nature of patient populations and clinical
practices. Traditional machine learning approaches have shown promise in addressing these challenges, but
their performance may degrade over time as the underlying data distribution changes, necessitating periodic
model retraining. Continual learning is a learning paradigm that enables models to adapt to new data continually
without forgetting previously learned knowledge. This approach is particularly well-suited to the early detection
of sepsis, as it allows models to capture the evolving nature of the condition, incorporate new diagnostic tools
and biomarkers, and provide personalized risk assessments. Despite its advantages, continual learning can
incur several safety issues in sepsis early detection. First, periodic retraining may lead to catastrophic forgetting,
which is a phenomenon that occurs when a machine learning model, upon excessive adaptation to new data
during retraining, loses its previously acquired knowledge from the original training data. Second, as models are
retrained with new data, there is a risk of overfitting to the most recent data, which may reduce their
generalizability to future patient populations or clinical settings. Ensuring a balance between updating the model
to adapt to changing conditions and maintaining generalizability is crucial for maintaining safety. The overarching
goal of this project is to develop safe continual learning methods for early sepsis detection that can adapt to new
data while avoiding catastrophic forgetting and maintaining generalization. First, we will develop methods that
update a small subset of weights in the existing model based on the new data, thereby preserving the previously
learned knowledge while adapting to the new information and leading to a more robust and adaptive model for
early sepsis detection. This will allow us to test the hypothesis that avoiding catastrophic forgetting in continual
learning will increase early prediction accuracy of sepsis for new patients without compromising accuracy for
historical patients. On several electronic health record datasets, we will perform experiments to test the
hypothesis that our methods will outperform state-of-the-art sepsis early prediction methods by more than 10%.
Second, we will develop methods to automatically detect distribution shift from streaming sepsis data, improve
the out-of-distribution (OOD) generalization capability of continual learning models, and quantify their prediction
uncertainty. This will allow us to test the hypothesis that improving OOD generalization in continual learning will
significantly improve the accuracy and robustness of early sepsis prediction on future patient cases.