INFORMING INTENSIVE CARE UNIT DIGITAL TWINS: DYNAMIC ASSESSMENT OF CARDIORESPIRATORY FAILURE TRAJECTORIES IN PATIENTS...

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Understanding sepsis patient clinical trajectories is vital for prognostication and resource allocation.
• Dynamic assessment of cardiorespiratory support is key to modeling patient progress.
• Existing models may not fully capture the diverse clinical pathways in intensive care units (ICUs).

Purpose of the Study:

• To identify common clinical trajectories for sepsis patients admitted to the ICU.
• To utilize electronic health record data for dynamic trajectory modeling.
• To inform the development of digital twin models for critical illness management.

Main Methods:

• Retrospective cohort study of 19,177 sepsis patients admitted to Mayo Clinic Hospitals ICUs over 8 years.
• Unsupervised machine learning (two-stage clustering) applied to model patient trajectories up to 14 days post-ICU admission.
• Modeling based on cardiorespiratory support dynamics and hospital discharge status.

Main Results:

• Four distinct clinical trajectories were identified: fast recovery (27%), slow recovery (62%), fast decline (4%), and delayed decline (7%).
• Mortality rates varied significantly: 3.5% for fast recovery, 3.6% for slow recovery, 99.7% for fast decline, and 97.9% for delayed decline.
• Trajectories were robustly distinguished by Charlson Comorbidity Index, APACHE III scores, and SOFA scores (P < 0.001).

Conclusions:

• Sepsis patient clinical trajectories can be effectively modeled using machine learning on EHR data.
• The identified trajectories offer a foundation for predictive models and digital twin decision support tools.
• These insights can improve shared decision-making and optimize resource planning in critical care.