---

Learning Objectives

By completing this assignment, you will:

• Design acceptance tests for a clinical AI model
• Implement basic monitoring and drift detection
• Create a governance checklist for AI deployment
• Think critically about what could go wrong in production

---

Background

You’ve built models. Now what? The gap between a working model and a safely deployed clinical tool is enormous. This assignment focuses on the “boring” but critical work of governance:

• Acceptance Testing: How do you validate a model works at your institution?
• Monitoring: How do you know when performance degrades?
• Governance: Who is responsible? What are the rules?

This is the stuff that determines whether your model helps patients or harms them.

---

Scenario

You’re a data scientist at a community hospital. The hospital wants to deploy a sepsis early warning model that predicts which patients in the general ward are at risk of developing sepsis in the next 6 hours.

The model was developed at a large academic medical center and published with impressive results (AUC 0.89). Your job is to:

1. Design local validation before deployment
2. Set up monitoring for after deployment
3. Create governance documentation

---

Instructions

Part 1: Acceptance Testing Design (30 points)

Before deploying any external model, you need to validate it works at your institution.

1.1 Define Your Test Cohort (10 pts)

Design a validation study:

• What patient population will you test on?
• What time period?
• How will you identify “ground truth” sepsis cases?
• What’s your minimum sample size and why?

Write this as a 1-paragraph protocol.

1.2 Choose Evaluation Metrics (10 pts)

Select 3-4 metrics that matter for this use case. For each:

• Why this metric?
• What threshold is acceptable?
• What would make you reject the model?

Consider: The model will generate alerts for nurses. False alarms cause alert fatigue. Missed sepsis cases can be fatal.

1.3 Acceptance Criteria Document (10 pts)

Create a formal acceptance criteria document with:

• Performance thresholds (with justification)
• Edge cases to test (e.g., post-surgical patients, pediatrics if applicable)
• Go/no-go decision criteria

---

Part 2: Monitoring Implementation (35 points)

Once deployed, you need to know if the model is still working.

2.1 Choose Monitoring Metrics (10 pts)

Select 2-3 statistics to monitor continuously. For each:

• What does it measure?
• What’s the expected baseline value?
• What change would trigger an alert?

Think about: input drift, prediction drift, outcome drift.

2.2 Implement Drift Detection (15 pts)

Write code that:

• Takes two distributions (reference and current)
• Calculates a drift statistic (e.g., PSI, KS test, chi-square)
• Returns whether drift is detected

Test your implementation with simulated data showing:

• No drift (distributions are similar)
• Drift detected (distributions differ)

2.3 Monitoring Dashboard Sketch (10 pts)

Create a simple visualization or sketch showing:

• Key metrics over time
• Alert thresholds
• Recent predictions summary

This can be a matplotlib figure or a hand-drawn sketch (photo) — the goal is to show what you’d want to see daily.

---

Part 3: Governance Framework (35 points)

Documentation that nobody reads is useless. Create governance artifacts that busy clinicians might actually use.

3.1 One-Page Model Card (15 pts)

Create a model card (max 1 page) including:

• Model purpose and intended use
• Training data summary
• Performance metrics
• Known limitations
• Who to contact with questions

3.2 Escalation Protocol (10 pts)

Define what happens when things go wrong:

• Who gets notified if drift is detected?
• What’s the threshold for pausing the model?
• Who makes the decision to turn it off?
• How do you communicate to clinical staff?

Write this as a simple flowchart or numbered steps.

3.3 Review Cadence (10 pts)

Propose a review schedule:

• How often do you review performance? (weekly? monthly?)
• What triggers an ad-hoc review?
• Who attends review meetings?
• What documentation is required?

---

Deliverables

File	Description
hw7_governance.py	Code for Part 2 (drift detection)
acceptance_criteria.md	Part 1.3 document
model_card.md	Part 3.1 one-page model card
escalation_protocol.md	Part 3.2 flowchart/steps
outputs/	Any figures (monitoring dashboard, etc.)

---

Grading Rubric

Component	Points
Part 1: Acceptance Testing	30
1.1 Test cohort design	10
1.2 Evaluation metrics	10
1.3 Acceptance criteria document	10
Part 2: Monitoring	35
2.1 Monitoring metrics	10
2.2 Drift detection implementation	15
2.3 Dashboard sketch	10
Part 3: Governance	35
3.1 Model card	15
3.2 Escalation protocol	10
3.3 Review cadence	10
Total	100

---

Resources

• Model Cards for Model Reporting (Mitchell et al.)
• FDA Guidance on AI/ML Medical Devices
• Hidden Technical Debt in ML Systems
• Population Stability Index (PSI) Explanation

---

Tips

• Think like a skeptic: What could go wrong? What would a critic ask?
• Be specific: “Monitor performance” is vague. “Review weekly AUC on patients over 65” is actionable.
• Keep it simple: A governance plan nobody follows is worse than none at all
• Consider your audience: Nurses, IT staff, and executives need different information
• This applies to your final project: Use this assignment as practice for your field guide