Tailored to Fit Sepsis Individuals: Medical Knowledge Aware Reinforcement Learning Model Offers Optimized Therapeutic...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Computational Biology

Background:

• Sepsis is a critical condition requiring timely and precise interventions.
• Reinforcement learning (RL) shows promise for sepsis treatment but struggles with patient heterogeneity and data imbalance.
• Existing RL models often fail to maintain optimal performance across diverse sepsis patient populations.

Purpose of the Study:

• To develop an advanced treatment recommendation model for sepsis patients with varying disease severity.
• To enhance the accuracy and effectiveness of therapeutic intervention strategies in sepsis management.
• To address the limitations of current RL models in handling heterogeneous patient data.

Main Methods:

• Proposed a novel model, Medical ontology knowledge and disease diagnosis Position aware Transfer learning - Double Dueling Deep Q Network (MPT-D3QN).
• Incorporated medical ontology knowledge and disease diagnosis positions using an attention mechanism for accurate patient state representation.
• Utilized an offline deep reinforcement learning algorithm and a transfer learning framework for training and cross-domain patient group analysis.

Main Results:

• The MPT-D3QN model achieved a higher expected return value (12.98 vs. 10.61) compared to current clinical practices on the MIMIC-III dataset.
• Demonstrated a significant reduction in estimated mortality from 13.20% to 6.15% across test datasets.
• Validated robust generalization capabilities on the external eICU dataset, outperforming state-of-the-art models.

Conclusions:

• The MPT-D3QN model provides optimal and interpretable treatment strategies for sepsis patients of varying severity.
• The model offers improved expected returns and reduced mortality, showcasing its clinical utility.
• This AI-driven approach enhances sepsis management by leveraging knowledge integration and transfer learning.