Back to Main Index Open Learning Roadmap

Deployment, Monitoring, and Model Drift in Health ML

A model is not finished when training ends. In health ML, the hard question is what happens after deployment: do predictions still match patients, does alert workload stay manageable, and when should a team recalibrate, retrain, or retire the model?

Beginner Learn what deployment means, why models drift, and why a good test score at launch is not a lifetime guarantee.
Intermediate Compare data drift, calibration drift, workflow drift, and alert burden while reading a simple monitoring dashboard.
Advanced Decide when to monitor only, recalibrate, refit, or retire a model based on changing prevalence, measurements, and care pathways.
4 Core monitoring questions: who changed, what changed, how much, and what action follows
Drift Performance can erode because patients, measurements, labels, or workflows no longer match training data
Action Most teams need a ladder: observe, recalibrate, refit, or retire

Why this page matters

The portal already covers safe splitting, class imbalance, calibration, and external validation. The next missing step is what students need for real-world health AI: once a model goes live, teams must monitor whether predictions remain trustworthy, clinically useful, and safe for the surrounding workflow.

Core idea: deployment is not the end of validation. It is the start of continuous checking under changing clinical reality.

Three kinds of drift every student should know

Beginner

Data or population drift

The incoming patients, prevalence, or feature values are no longer similar to the development data.

  • Examples: younger patients, different admission patterns, new referral sources.
  • Alert volume and positive predictive value can change even if the code is unchanged.
Intermediate

Calibration drift

The ranking may still look acceptable, but predicted probabilities become too high or too low.

  • Often appears after prevalence changes or workflow changes.
  • Important when risk estimates trigger counseling, treatment, or escalation.
Advanced

Workflow or concept drift

The meaning of the outcome or the use of the prediction changes because practice changed.

  • Examples: new treatment guidelines, new triage rules, different documentation patterns.
  • These settings may require more than recalibration.

Interactive monitoring lab: the same model can age in different ways

Select a live deployment scenario. Watch how discrimination, calibration, alert burden, and recommended action move together. This is a simplified monitoring dashboard, but it teaches the right habit: do not follow one metric in isolation.

Deployment monitor

Stable first month
0.84 Discrimination (AUC)
Good Calibration at current threshold
12% Patients triggering alerts
Monitor Recommended response

Signal strength

Population shift Low
Feature measurement shift Low
Workflow impact Low

What to notice

The launch setting still resembles the development data. The goal is to establish a baseline before interpreting later changes.

  • Track the same metrics on a fixed schedule.
  • Record workload, not just accuracy.
  • Keep thresholds fixed unless governance approves a change.
Monitoring warning: a stable AUC does not guarantee safe deployment. Calibration or alert burden can break first.

Possible responses when monitoring detects a problem

Monitor only

Use this when shifts are small and performance remains clinically acceptable.

  • Good for early baseline building.
  • Requires predefined review intervals.

Recalibrate

Use this when ranking still works, but probabilities are systematically too high or too low.

  • Typical after prevalence or case-mix changes.
  • Must be documented and rechecked.

Refit or update

Use this when feature relationships or measurements changed enough that calibration alone is not enough.

  • May require new data collection and validation.
  • Thresholds may need redesign too.

Retire or pause

Use this when the workflow changed so much that the model no longer matches the intended decision.

  • Sometimes the safest action is to stop using the tool.
  • Retirement is a valid lifecycle outcome.

A practical monitoring workflow for health ML

1

Freeze the intended use

Define who the model is for, when it runs, who sees the output, and what action follows.

2

Choose a baseline window

Measure early deployment performance and workload before deciding what counts as abnormal.

3

Monitor several signals

Track discrimination, calibration, prevalence, missingness, alert rate, and downstream workload together.

4

Investigate the cause

Ask whether the change came from patients, measurements, labels, or workflow before changing the model.

5

Use a response ladder

Escalate from monitoring to recalibration, update, or retirement based on evidence and governance.

6

Document every change

Version the model, thresholds, and dates of updates so future evaluations remain interpretable.

Common mistakes

Watching only AUC

A model may keep decent ranking while predicted risks or alert workload become unreliable.

Changing thresholds informally

Threshold changes alter workload and safety. They should be governed, not improvised during daily operations.

Ignoring label changes

If documentation or adjudication changes, observed performance may shift even without a real biological change.

Assuming deployment is permanent

Every model should have criteria for update, pause, or retirement.

Safety note: in healthcare, false alarms create workload and alert fatigue, while missed cases create harm. Monitoring must respect both.

Quick self-check

Use these short cases to confirm the main deployment logic before moving on.

Case 1

A deployed model keeps similar AUC, but predicted risks are now too high for most patients. What is the most likely issue?

Correct answer: Calibration drift. Ranking can stay acceptable while the probability scale becomes unreliable.

Case 2

A hospital adopts a new lab analyzer and several model inputs now shift abruptly. What should you suspect first?

Correct answer: Measurement drift. Input distributions may change because the measurement system changed.

Case 3

Which response is strongest when the clinical workflow itself changed and the model no longer matches the decision point?

Correct answer: Consider refit or retirement. A workflow mismatch may be deeper than a probability problem.

How this connects to other portal pages

After external validation

Use external validation to decide whether a model travels before it goes live.

With calibration and thresholds

Use the calibration guide to understand which deployment changes require recalibration or a threshold review.

With class imbalance

Read the class imbalance guide when prevalence shift changes precision and alert burden.

Before future lifecycle topics

This page prepares students for deeper topics such as model updating, governance, impact studies, and reporting standards.