Pruning the Path to Optimal Care: Identifying Systematically Suboptimal Medical Decision-Making with Inverse...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Clinical decision analysis

Background:

• Observational clinical data contains complex decision-making patterns.
• Identifying suboptimal clinician actions is crucial for improving patient outcomes.
• Existing methods may struggle to differentiate optimal from suboptimal behaviors in complex datasets.

Purpose of the Study:

• To apply Inverse Reinforcement Learning (IRL) for analyzing clinician decision-making in intensive care units (ICUs).
• To identify and characterize suboptimal clinician actions by comparing them to peer behavior.
• To uncover underlying clinical priorities and values from real-world ICU data.

Main Methods:

• A two-stage IRL approach was developed.
• An intermediate step was implemented to prune non-consensus trajectories.
• The model was applied to observational ICU data encompassing diverse patient populations and conditions.

Main Results:

• The IRL model successfully identified suboptimal clinician actions.
• Pruning non-consensus behaviors revealed distinct clinical priorities.
• The impact of removing suboptimal actions differed across various diseases.
• Certain demographic groups were disproportionately affected by the benefits of action optimization.

Conclusions:

• Novel IRL application effectively models clinical decision-making and identifies deviations from consensus.
• The findings highlight the heterogeneity in clinical practice and its impact on different patient groups.
• This approach offers a pathway to refine clinical guidelines and improve healthcare quality by understanding decision-making nuances.