CBS Faculty Live: Leveraging AI to Improve Healthcare with Professor Carri Chan

Key Concepts

• Predictive Analytics: Using historical data and statistical modeling to forecast future events (e.g., patient deterioration, staffing needs).
• Proactive vs. Reactive Care: Shifting from treating patients only after they become critically ill to intervening early based on risk scores.
• Early Warning Scores (EWS): Algorithmic tools that monitor patient vitals and trends to predict health deterioration.
• Capacity Constraints: The operational reality that resources (ICU beds, staff, colonoscopy slots) are limited, necessitating optimized allocation policies.
• Surge Staffing: A workforce management strategy that adjusts staffing levels in real-time based on predicted demand to avoid burnout and maintain quality.
• Cannibalization: A phenomenon where low-risk patients seeking care "crowd out" high-risk patients in a capacity-constrained system.
• Area Under the Curve (AUC): A metric for predictive accuracy; however, the speaker argues it is insufficient for operational decision-making when capacity is limited.

1. The State of AI in Healthcare

• Data Proliferation: Hospitals generate approximately 50 petabytes of data annually, yet roughly 97% remains unused, often due to unstructured formats (clinical notes) and interoperability challenges.
• Economic Impact: Healthcare accounts for 18% of U.S. GDP, with an estimated $700–$900 billion in annual waste. Widespread AI adoption could potentially save $365 billion.
• Regime Change: The industry is transitioning from reactive, acute-care models to proactive, data-driven systems.

2. Clinical Applications: Proactive Care

• Advanced Alert Monitor (AAM): Developed with Kaiser Permanente, this model analyzes 24 hours of patient data (vitals, labs, trends) to predict deterioration within the next 12 hours.
  • Outcome: Implementation led to a 31% reduction in 30-day mortality and a 7-hour reduction in ICU length of stay.
  • Methodology: A 10-year process involving initial evidence gathering, a "silent" phase (monitoring without alerts), and a phased rollout across 21 hospitals.
• Preventive Screening: A project with Geisinger Health Systems used predictive modeling to identify patients overdue for colonoscopies.
  • Outcome: Targeted nursing outreach increased screening uptake by 7% and resulted in a 6.2% reduction in 2-year mortality.

3. Operational Applications: Workforce Management

• Emergency Department (ED) Staffing: Hospitals face high demand uncertainty. The speaker’s research uses real-time data (Google search trends, weather, current ED volume) to optimize staffing.
• Base vs. Surge Staffing:
  • Base Staffing: Set weeks in advance; lower cost but higher uncertainty.
  • Surge Staffing: Real-time adjustments (overtime, agency nurses); higher cost but necessary for spikes.
• Case Study: A pilot at Hackensack Meridian Health used predictive algorithms to guide nursing managers, resulting in an estimated $1.4 million in savings without compromising patient wait times or quality.

4. Frameworks for Operationalizing AI

• The "Capacity-Aware" Approach: The speaker argues that traditional AI development focuses too heavily on predictive accuracy (AUC). In practice, if a system is capacity-constrained, a model with lower AUC but higher precision at low-false-positive rates is often superior.
• The "Cannibalization" Risk: Increasing the number of flagged patients can lead to "independent requests" (patients seeking care on their own) that crowd out the high-risk patients the model was intended to prioritize.

5. Key Arguments and Perspectives

• The "No-Harm" Constraint: Unlike other industries, healthcare cannot use "exploration" (trial and error) in AI because of the ethical risks to patient safety.
• Data Representation: AI models are only as good as their training data. If datasets lack representation from marginalized communities, the resulting algorithms may inadvertently widen health disparities.
• Structural Barriers: The fee-for-service payment model acts as a disincentive for proactive care, as hospitals often lose revenue when patients stay healthy or avoid hospitalizations.

6. Notable Quotes

• "If you don't think about the types of decisions you're going to base on these [models] and how you're going to actually operationalize it, you're never going to be able to realize these benefits." — Professor Carrie Chan
• "AI models are only as good as the data that they are trained on." — Professor Carrie Chan

7. Synthesis and Conclusion

The transition to a proactive healthcare future requires more than just high-accuracy algorithms; it requires a deep integration of predictive modeling with operational reality. Success depends on aligning AI outputs with capacity constraints, ensuring data representativeness to prevent bias, and overcoming the financial disincentives inherent in the fee-for-service model. The most effective implementations are those that treat AI as a tool for decision support rather than a standalone solution, requiring long-term commitment and close collaboration between data scientists, clinicians, and administrators.