Dynamic selection of models for a ventilator-management advisor

Area of Science:

• Biomedical Engineering
• Computational Physiology
• Medical Informatics

Background:

• Mechanical ventilation requires precise management, often aided by computer programs called ventilator-management advisors (VMAs).
• VMAs use patient-specific physiologic models to interpret data and predict outcomes of ventilator setting changes.
• Selecting an appropriate model complexity is crucial to balance accuracy and computational speed in time-critical clinical scenarios.

Purpose of the Study:

• To present a novel method for selecting physiologic models that optimize the trade-off between prediction accuracy and computation-time complexity for VMAs.
• To address the challenge of choosing models that are sufficiently detailed yet computationally feasible for real-time decision-making.

Main Methods:

• A local search algorithm within a graph of models (GoM) is employed to identify optimal models.
• Belief networks are used to compute the probability of model adequacy based on prior information.
• Goodness-of-fit measures assess the conditional probability of model adequacy given observed patient data.

Main Results:

• The method effectively balances prediction accuracy and computation-time complexity in model selection for VMAs.
• A graph of physiologic models, ranging from simple (VentPlan) to complex (VentSim), was implemented and utilized.
• The approach demonstrated suitability for time-constrained decision tasks in mechanical ventilation management.

Conclusions:

• The presented method provides an effective strategy for selecting patient-specific physiologic models for ventilator-management advisors.
• Optimizing model selection enhances the performance of VMAs in critical care settings.
• This approach facilitates better-informed clinical decisions during mechanical ventilation.