Lecture Notes on AI in Clinical Deterioration Prediction
Introduction
- Speaker Andrew's background:
- Emergency internal medicine and critical care
- Transitioned into data science
- Developed AI models with colleagues Marg and Jan
Project Overview
- Focus on clinical deterioration in patient wards
- Aim to prevent escalations (ICU transfer, rapid response activation, code blue, mortality)
- Excluded populations: pregnant patients, mental illness, individuals under 18, observed wards
Project Progress
- Began with "failure to rescue"
- Evolved into "nurturing intuition score"
- Currently using a sophisticated AI model named 'Nydia'
AI Model Goals
- Predict clinical deterioration using Electronic Health Records (EHR)
- Intervene before ICU escalations by predicting likelihood
- Compare AI predictions with clinical intuition
Validation Approach
- Emphasize clinical validity of the AI model
- Conduct chart reviews:
- Patients who did not deteriorate
- Patients who deteriorated
- Review data from rapid response teams
- Consider cognitive decline and vascular issues
Challenges and Considerations
- No standard for clinical validation of AI models
- Goal to use clinicians as the gold standard, not AI
- If AI model becomes outdated, retraining may be necessary
- Evaluate against other models (NEWS, MUSE, Epic deterioration index)
Discussion on Model Integration
- Potential of AI models to reduce fatigue and nurse burnout
- AI as a predictive measure to prevent escalated outcomes
- Opportunity to render rapid response team unnecessary
Implementation and Clinical Workflow
- Integrate AI model silently until validation is complete
- Considerations for production in the Epic system
- Discussions on clinical protocols for predicted deterioration
Additional Considerations and Next Steps
- Importance of detailed chart reviews
- Use REDCap for data collection
- Collaborate with specialists for setup
- IRB submission process for research
Summary
- Value-based care initiatives included
- Potential for publishing findings and integrating AI into clinical practice
- Discussions on handling AI lifecycle and improvements
References
- Related papers and models for further reading
This lecture provides a comprehensive overview of the development, validation, and potential of an AI model designed to predict patient deterioration in clinical settings. It suggests ultimate outcomes for reducing burdens on healthcare systems and nurses, while also comparing existing models.