CPV Drift Detection Sensitivity (CPVDDS)

GMI-PH-004Pharmaceutical OperationsInfrastructure-Dependent

Definition

The smallest process drift rate the continued process verification model reliably detects, defined as the minimum drift at which the model alerts at least 80 percent of the time. Expresses detection capability as a sensitivity floor rather than an accuracy percentage.

The Measurement Gap

Traditional process capability indices like Cpk are calculated periodically, often annually during the product quality review. They are snapshots, not continuous monitors. A Cpk calculated at the start of a slow drift shows excellent capability; a Cpk calculated at the end shows the process at the edge. Between those two calculation events, dozens of batches may ship from a process slowly leaving its validated state. CPV models designed around control chart rules detect step-changes (sudden shifts in a parameter) but slow, directional drift that consumes the validated range at fractions of a percent per batch is structurally invisible. Each individual data point looks normal. The model evaluates trees one at a time and never sees the forest trending toward the boundary.

Decision Supported:

Can our CPV model actually detect slow drift, or does it only catch step-changes? If it can only detect step-changes, we have a blind spot for gradual process degradation.

Formula

Notation:

CPVDDS = min(δ) such that P(Alert | drift rate = δ) ≥ 0.80

Variables:

  • δ = drift rate as percentage of validated range consumed per time unit (e.g., 0.5% per batch or 1% per week)
  • P(Alert | drift rate = δ) = probability the CPV model generates an alert given the injected drift rate
  • Validated range = the approved operating range for the parameter under monitoring

Method: Inject known synthetic drift signals of varying magnitudes into historical CPV data. Measure the smallest drift rate the model consistently detects (alerting in more than 80% of injected cases) before the simulated process crosses a validated boundary.

Worked Example:

Cell culture pH with validated range 6.95–7.15 (range = 0.20 pH units), historical noise σ = 0.015 pH units. Injecting synthetic drift at 0.5% per batch: 12% alert rate. At 1.0%: 31%. At 2.0%: 67%. At 3.0%: 84%, exceeding the 80% threshold. CPVDDS = 3.0% of validated range per batch. At that rate, mid-range to boundary drift takes approximately 17 batches.

Sign convention:

Lower CPVDDS is better. A value of 1% means the model detects very slow drift. A value of 5% means only moderate drift is visible.

Edge Cases:

  • Test both upward and downward drift as models may be asymmetrically sensitive
  • Report CPVDDS alongside process noise estimate; 3% in a low-noise process differs significantly from 3% in a high-noise process
  • Assess CPVDDS per parameter if the CPV model uses different algorithms per parameter

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher
Reporting Frequency
Continuous; monthly drift review
Category
AI Model Fidelity & Drift
← Batch Record Auto-Clear Accuracy (BRACA)Inspection Sensitivity Distribution Index (INSDI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment