Comorbidity network analysis using graphical models for electronic health records.

Area of Science:

• Critical care medicine
• Computational biology
• Medical informatics

Background:

• Understanding comorbidity networks in critical care units (CCUs) is crucial for timely diagnosis and improved patient outcomes.
• Existing methods for analyzing disease relationships may not fully capture complex comorbidity patterns.

Purpose of the Study:

• To identify the comorbidity network among CCU patients using a novel machine learning (graphical modeling) method.
• To compare the performance of the machine learning method against traditional pairwise methods in simulations.

Main Methods:

• A cross-sectional study utilizing electronic health record data from the Medical Information Mart for the Intensive Care-3 (MIMIC-3) dataset (2001-2012).
• Application of a graphical modeling method to analyze 654 diagnosis categories among 46,511 CCU patients.
• Validation of identified associations by medical professionals and comparison with traditional pairwise methods via simulation.

Main Results:

• The graphical modeling method identified 2,806 associations among 510 diagnosis categories, with expert validation confirming medical consistency.
• Key findings include strong associations like 'poisoning by psychotropic agents' and 'accidental poisoning by tranquilizers' (logOR 8.16).
• The diagnosis 'disorders of fluid, electrolyte, and acid-base balance' was the most connected (63 associations), and the method outperformed traditional approaches in simulations, partitioning diagnoses into 14 modularity classes.

Conclusions:

• The developed graphical modeling method effectively infers a clinically relevant comorbidity network in CCU patients.
• This data-driven approach demonstrates superior performance compared to traditional methods in simulation studies.
• The identified comorbidity network has the potential to significantly assist clinicians in faster diagnosis and reducing missed diagnoses in CCU settings.