CoAI

CoAI enables cost-aware predictive modeling in healthcare by selecting minimal sets of low-cost patient features to produce accurate predictions under predefined budget constraints.

Key Features:

• Cost Efficiency: Selects and prioritizes low-cost, efficient patient features to minimize data acquisition effort and expense.
• Model Agnosticism: Supports training with deep neural networks and tree ensemble models for predictive tasks.
• Axiomatic Feature Attribution: Employs axiomatic feature attribution methods to estimate feature importance for selection.
• Budget-Conscious Optimization: Identifies high-performance models within a predefined budget without requiring manual tuning of the cost-versus-performance tradeoff.

Scientific Applications:

• Prehospital acute traumatic coagulopathy prediction: Improves prediction of prehospital acute traumatic coagulopathy using selected low-cost features.
• Intensive care mortality prediction: Enhances prediction of intensive care unit mortality with cost-constrained feature sets.
• Outpatient mortality prediction: Improves outpatient mortality prediction while reducing data-gathering requirements.
• Population-scale trauma surveillance (United States): Applied across all trauma patients in the United States, could alert providers to tens of thousands more dangerous events than current methods while reducing data-gathering time by approximately 90% and saving an estimated 200,000 cumulative hours per year.

Methodology:

Uses cost-aware feature selection and budget-constrained optimization, employs axiomatic feature attribution to estimate feature importance, and trains predictive models using deep neural networks or tree ensemble models without requiring manual tuning of the cost-versus-performance tradeoff.