OnAI-Comp: An Online AI Experts Competing Framework for Early Sepsis Detection

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Sepsis poses significant public health challenges due to high mortality, morbidity, and economic burden.
• Existing early sepsis prediction models often use offline approaches, limiting their ability to dynamically adapt to growing patient datasets.
• Retraining offline models with new data is computationally intensive, hindering real-time clinical application.

Purpose of the Study:

• To develop a novel online learning framework for dynamic early sepsis detection.
• To address the limitations of computationally expensive retraining in offline sepsis prediction models.
• To improve the adaptability and efficiency of sepsis prediction systems in real-world clinical settings.

Main Methods:

• Proposed an Online Artificial Intelligence Experts Competing Framework (OnAI-Comp) utilizing a Multi-armed Bandit online learning algorithm.
• Selected diverse machine learning models as 'artificial intelligence experts' within the framework.
• Evaluated model performance using average regret to assess convergence to optimal prediction strategies.

Main Results:

• Experimental analysis confirmed that the OnAI-Comp framework converges to an optimal prediction strategy over time.
• The online learning approach effectively incorporates new patient data for continuous model improvement.
• The framework supports clinically interpretable predictions via existing local interpretable model-agnostic explanation technologies.

Conclusions:

• The OnAI-Comp framework offers a computationally efficient and dynamically adaptable solution for early sepsis detection.
• Online learning with Multi-armed Bandits provides a robust method for improving sepsis prediction models with real-time data.
• Clinically interpretable predictions can aid healthcare professionals in timely decision-making, potentially improving patient survival rates.