Governing Decisions of Probability Cutoffs in Clinical AI Deployment: A Case Study of Asthma Exacerbation Prediction

This paper argues that selecting probability cutoffs for clinical AI models, such as asthma exacerbation predictors, should be treated as a structured governance process that balances statistical performance with operational feasibility and stakeholder values, rather than relying solely on technical optimization.

The Gist

Imagine you are the captain of a massive ship (a hospital) trying to navigate through a foggy sea (patient care). You have a high-tech radar (an AI model) that can detect icebergs (asthma attacks) before they happen. But here's the catch: the radar doesn't just say "Iceberg!" or "No Iceberg." It gives you a probability score, like a dial that spins from 0% to 100%.

The big question is: At what number on that dial do you sound the alarm and tell the crew to start bailing water?

This paper is about how a team at Mayo Clinic figured out exactly where to set that alarm dial for predicting asthma attacks in kids, and why simply picking the "mathematically perfect" number is a bad idea.

The Problem: The "Perfect" Number is a Trap

In the world of math and statistics, there are formulas to find the "optimal" cutoff. It's like trying to find the perfect temperature for a shower. Math might say, "The perfect temperature is 98.6°F."

But in real life, if you set your shower to that exact number, you might get scalded if the water pressure drops for a second, or you might get cold if the heater lags. You need a little wiggle room.

Similarly, the authors found that if you just let the computer pick the "best" statistical number to predict asthma attacks, you might end up with two disasters:

1. The False Alarm Flood: The alarm goes off for almost everyone. The nurses and doctors get so many alerts that they stop listening to them (this is called "alert fatigue"). It's like a fire alarm that goes off every time someone burns toast; eventually, nobody runs when the real fire starts.
2. The Missed Danger: You set the alarm so high that you only catch the biggest fires, but you miss the small ones that could still burn the house down.

The Solution: A Team Huddle (Governance)

Instead of letting a computer decide alone, the doctors and data scientists held a "town hall meeting." They treated the decision not as a math problem, but as a policy decision.

Here is how they did it, using a simple analogy:

The Analogy: The Security Gate
Imagine the hospital is a stadium, and the AI is a security scanner at the gate.

• Sensitivity (Catching the bad guys): If you set the scanner to be super sensitive, it will catch every single person with a tiny pocket knife, but it will also stop people with metal belt buckles, keys, and even a spoon. The line gets backed up for hours.
• Specificity (Letting the good guys through): If you set the scanner to be very strict, only people with guns get stopped. But you might miss the guy with a small knife, and he gets into the stadium.

The team had to decide: How much traffic jam are we willing to tolerate to make sure we don't miss a dangerous person?

The Process: Turning Math into Real Life

The researchers didn't just show the doctors a graph. They translated the math into real-world consequences:

• The Math: "If we pick Threshold A, we have 90% sensitivity."
• The Translation: "If we pick Threshold A, our 167 doctors will have to check 1,103 patients this year. That's about 7 extra patients per doctor every single day. Can they handle that?"
• The Math: "If we pick Threshold B, we have 60% sensitivity."
• The Translation: "If we pick Threshold B, we will only check 390 patients, but we will miss 125 kids who might have a severe asthma attack."

The Decision: Finding the "Goldilocks" Zone

The doctors realized that missing an asthma attack (a false negative) is very scary and dangerous. However, checking a kid who doesn't need it (a false positive) is just a little bit of extra paperwork and a phone call.

So, they decided to lean towards catching more kids, even if it meant a few more phone calls. They picked a threshold that:

1. Caught about 87% of the kids who would have an attack.
2. Created a workload that was manageable (about 61% of the patients flagged), meaning doctors wouldn't be overwhelmed.

They called this the "Goldilocks" threshold: not too strict, not too loose, but just right for the team's capacity.

The Big Takeaway: Write It Down!

The most important part of this paper is the Governance Template.

The authors realized that in the past, these decisions were often made in the back of a room with no notes. "Hey, let's just use this number."

They created a standardized report card (like a recipe) that anyone can read later. It documents:

• What numbers we looked at.
• Why we picked the one we did.
• How many extra patients the doctors will have to see.
• What happens if the system changes later.

In a Nutshell

This paper teaches us that AI isn't just code; it's a tool for people.

Setting the rules for an AI isn't about finding the perfect mathematical answer. It's about having a conversation between the data scientists and the doctors to ask: "How much work can our team handle? How much risk are we willing to take? And how do we write this down so we don't forget why we made this choice?"

It turns a cold, hard number into a thoughtful, human decision that keeps patients safe without burning out the staff.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in the deployment of Clinical Artificial Intelligence (AI): the selection of operational probability cutoffs (thresholds).

• The Technical Challenge: AI models output continuous probability scores, but clinical workflows require discrete binary decisions (e.g., "alert" vs. "no alert").
• The Limitation of Current Practices: Threshold selection is often treated as a purely statistical optimization problem (e.g., maximizing F1-score, Youden's index, or ROC curve proximity). This approach assumes equal costs for false positives and false negatives and ignores real-world constraints.
• The Consequence: Statistically "optimal" thresholds often lead to operationally infeasible alert volumes (causing alert fatigue) or unacceptably high rates of missed adverse events.
• The Governance Gap: There is a lack of structured, documented governance processes for justifying threshold choices, despite regulatory bodies (FDA, NIST) emphasizing the need for clinical justification and performance documentation at specific thresholds.

2. Methodology

The authors propose and implement a governance-informed, three-step framework for threshold selection, applied to a pediatric asthma exacerbation (AE) prediction model.

A. The Three-Step Framework

1. Candidate Threshold Identification:
   • Selected multiple points along the Receiver Operating Characteristic (ROC) curve, including:
     • Statistical maximum F1-score.
     • Youden's index (maximizing sensitivity + specificity - 1).
     • Neighboring points illustrating trade-offs between sensitivity, Positive Predictive Value (PPV), and flag rates.
2. Structured Clinical Deliberation:
   • Conducted a facilitated session with an advisory panel of four practicing clinicians.
   • Contextualization: Translated statistical metrics into operational case counts using local data:
     • Cohort: 1,235 pediatric asthma patients.
     • Prevalence: 24.1% (298 exacerbation events).
     • Provider Capacity: 167 providers (approx. 2 asthma patients/provider/month).
   • Evaluation: Panelists reviewed projected annual case counts for captured events, missed events, and total flagged patients for each candidate threshold.
3. Governance Documentation:
   • Utilized a standardized template to document the rationale, performance metrics, workflow implications, and monitoring plans for the final decision.

B. Data & Metrics

• Model Output: Continuous risk score for AE within 12 months.
• Key Metrics Evaluated: Sensitivity, PPV, Flag Rate, and projected workload (number of patients requiring follow-up).
• Intervention Context: The triggered intervention was low-risk (chart review and patient outreach), influencing the panel's tolerance for false positives.

3. Key Results

The study evaluated five candidate thresholds, revealing significant divergence between statistical optimality and operational feasibility.

• Performance Trade-offs:
  • Sensitivity Range: 58.1% to 97.1%.
  • PPV Range: 44.3% to 26.1%.
  • Flag Rate Range: 31.6% to 89.2% of the patient population.
• Operational Impact:
  • High-Sensitivity Scenario (Low Threshold): Captured 289/298 events but flagged 1,103 patients (89.2%), overwhelming provider capacity.
  • Statistical Optimum (Max F1): Captured 173/298 events but flagged 390 patients (31.6%).
  • Selected Consensus (Threshold 3): Captured 258/298 events (86.8% sensitivity) while flagging 756 patients (61.3%).
    • Rationale: This workload was deemed manageable (approx. one alert per pediatrician per month) while capturing the majority of preventable events.
• Deliberation Outcomes:
  • Clinicians prioritized missed event reduction (false negatives) over alert volume due to the life-threatening nature of asthma exacerbations.
  • However, they rejected the highest sensitivity threshold due to the risk of alert fatigue and dilution of attention.
  • The final decision balanced high detection rates with the physical limits of the care team.

4. Key Contributions

1. Governance Framework: Proposes a structured, reproducible process for threshold selection that moves beyond statistical optimization to include clinical utility, workflow feasibility, and stakeholder input.
2. Operational Translation: Demonstrates the necessity of translating abstract metrics (AUC, F1) into concrete operational projections (e.g., "number of alerts per provider per month") to facilitate informed decision-making.
3. Standardized Documentation Tool: Introduces a governance documentation template (Table 1) that captures:
   • Candidate thresholds and rationale.
   • Expected alert rates and sensitivity/PPV.
   • Workflow integration details.
   • Monitoring plans.
4. Shift in Paradigm: Argues that threshold selection is an organizational governance process involving value judgments on harm tolerance and resource allocation, rather than a purely technical task.

5. Significance and Implications

• Clinical Safety: Prevents the deployment of models that either miss critical events or overwhelm clinicians with noise, thereby protecting patient safety and provider well-being.
• Regulatory Compliance: Provides a mechanism to meet emerging regulatory requirements (FDA, Coalition for Health AI) for documenting model performance and justification at specific operating points.
• Scalability: The framework is generalizable to other predictive models and clinical settings, offering a blueprint for transparent AI governance.
• Transparency: By explicitly documenting the trade-offs and value judgments made during threshold selection, health systems can build trust in AI systems and facilitate future audits or model updates.

Conclusion: The paper concludes that treating threshold selection as a governance process, supported by structured deliberation and standardized documentation, is essential for the responsible, safe, and effective deployment of clinical AI.