ADT²R: Adaptive Decision Transformer for Dynamic Treatment Regimes in Sepsis.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Computational Biology

Background:

• Dynamic treatment regimes (DTRs) are crucial for managing complex diseases like sepsis, with current methods often relying on offline reinforcement learning (RL) from electronic health records.
• Existing sepsis DTR studies face limitations, including divergence from clinician decisions, and failure to account for patient heterogeneity, short-term health state transitions, and the patient state-prescription relationship.

Purpose of the Study:

• To propose a novel framework, the adaptive decision transformer for DTR (ADT2R), for optimizing sepsis treatment recommendations.
• To address limitations of existing methods by incorporating patient heterogeneity and dynamic health state information into treatment decision-making.

Main Methods:

• Developed ADT2R, a framework utilizing a trajectory-optimization module trained with supervised learning for treatment recommendations.
• Employed multihead self-attention to capture time-varying patterns in sepsis patients and an actor-critic (AC) algorithm to estimate and incorporate short-term health state changes.
• Validated the framework on the Medical Information Mart for Intensive Care III (MIMIC-III) dataset.

Main Results:

• The ADT2R framework demonstrated effectiveness in recommending optimal treatment actions at each time step.
• Performance was comparable to state-of-the-art methods on the MIMIC-III dataset, indicating its clinical relevance.
• The model successfully integrated patient heterogeneity and evolving health states into treatment decisions.

Conclusions:

• ADT2R offers a promising approach for personalized sepsis management by adaptively recommending treatments.
• The framework enhances DTRs by considering individual patient characteristics and dynamic health trajectories.
• This work advances the application of AI in critical care, particularly for complex conditions like sepsis.