Real-World Case Studies: AI Improving Patient Outcomes
From the course AI for Nurses
Three leading health systems show what AI-assisted nursing looks like in practice — with measurable results.
Case Study 1: Johns Hopkins — Sepsis Early Warning
The Problem: Sepsis kills approximately 270,000 Americans annually, often because the 6-hour treatment window is missed on busy floors.
The AI Solution: Johns Hopkins deployed the Targeted Real-time Early Warning System (TREWS), an ML model that continuously analyzes vital signs, lab values, and nursing notes to flag sepsis risk.
Nursing Role: Nurses receive alerts in the EHR and confirm or dismiss within 1 hour. The model learns from dismissal patterns to reduce false positives.
Outcomes:
- 18 % reduction in sepsis mortality in the first 2 years
- Median alert-to-antibiotic time dropped from 3.8 hours to 1.7 hours
- False positive rate fell 40 % after nurses trained the model feedback loop
Key lesson: Nurse feedback on false positives was essential to the model's improvement — AI without nursing input degrades over time.
Case Study 2: UC San Diego Health — AI Documentation Assistant
The Problem: ICU nurses averaged 35 % of shift time on documentation, leaving less time for direct patient care.
The AI Solution: UCSD piloted an ambient AI documentation tool that listens to nurse-patient interactions (with consent), auto-generates structured nursing notes, and flags documentation gaps.
Outcomes:
- Documentation time fell from 35 % to 22 % of shift time
- Nurse satisfaction scores increased 28 points on the unit
- Compliance with mandatory nursing assessment fields improved from 71 % to 94 %
- No HIPAA incidents in the 18-month pilot
Key lesson: Nurses needed 2 weeks to trust the auto-generated notes. Training time on reviewing rather than re-writing AI content was critical.
Case Study 3: Vanderbilt University Medical Center — Medication Safety AI
The Problem: Traditional medication alert systems had a 95 % override rate — nurses learned to ignore them.
The AI Solution: Vanderbilt implemented an ML-based alert system that ranks alerts by actual risk score rather than triggering on any interaction. Low-risk interactions are suppressed; high-risk ones include a plain-language explanation.
Outcomes:
- Alert volume reduced by 54 %
- Override rate for high-priority alerts fell from 52 % to 9 %
- Near-miss medication events caught increased by 31 % (fewer alerts, more attention per alert)
- Nurse alert fatigue scores (validated instrument) improved by 4.2 points
Key lesson: Fewer, smarter alerts outperform comprehensive alerts. Nurses cannot maintain vigilance when every interaction triggers a warning.
Common Threads Across All Three Cases
- Nurse involvement in design — all three systems had bedside nurses on the design team
- Feedback loops — nurses could correct the AI, improving it over time
- Transparent explanations — alerts included the reason, not just the flag
- Phased rollout — pilot units before hospital-wide deployment revealed workflow gaps
- Measured outcomes — success was defined in patient outcomes, not just adoption rates
Key Takeaways
- •Johns Hopkins TREWS cut sepsis mortality 18 % by alerting nurses within the 6-hour treatment window
- •UCSD's ambient documentation assistant freed 13 % of shift time for direct patient care
- •Vanderbilt's AI cut alert volume 54 %, making remaining alerts 3x more likely to be acted on
- •Nurse involvement in design and feedback loops are the consistent success factor across all three sites
Progress is kept in this browser only.