Inspection Sensitivity Distribution Index (INSDI)

GMI-PH-002Pharmaceutical OperationsIntegration-Required

Definition

Divergence between the current distribution of detected defect categories and the validated baseline distribution. Catches a vision model losing sensitivity to safety-critical defects while gaining it on cosmetic ones, even when the overall reject rate holds steady.

The Measurement Gap

Overall Reject Rate (the metric everyone watches) is blind to compensatory sensitivity shifts because it tracks the total, not the composition. A computer vision inspection model can quietly lose sensitivity to safety-critical defects (particulate matter in injectables) while gaining sensitivity to cosmetic defects (micro-scratches), with the net reject rate remaining stable. OEE sees no change, Batch Success Rate shows no decline, and even individual defect-category reject rates may stay within their statistical process control limits. The problem is the relative distribution shift: the model's sensitivity profile has drifted from its validated baseline in a way that no single-metric dashboard can detect. Environmental changes such as lighting upgrades, camera aging, or formulation changes drive these shifts without triggering any traditional alarm.

Decision Supported:

Is our vision system still catching the same types of defects it was validated to catch, or has its sensitivity profile shifted so that it is over-detecting some defects and under-detecting others?

Formula

Plain language: For each defect category, calculate the current reject rate as a fraction of total rejects. Compare this distribution to the validated baseline distribution using the Jensen-Shannon divergence, a symmetric measure of how different two probability distributions are. Subtract from 1 so that the result is 1.0 when distributions are identical and approaches 0 when completely different.

Notation:

ISDI = 1 − JSD(P_current || P_baseline)

Where JSD(P || Q) = ½ KL(P || M) + ½ KL(Q || M), M = ½(P + Q)
KL(P || M) = Σ P(i) × ln(P(i) / M(i))

Variables:

Variable Definition Unit
P_current(i) Fraction of total rejects in defect category i over the current measurement period Ratio
P_baseline(i) Fraction of total rejects in defect category i from the validated baseline period Ratio
i Defect categories as defined in the inspection validation protocol Enumeration

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher
Reporting Frequency
Monthly; stratified by defect class
Category
AI Model Fidelity & Drift
← CPV Drift Detection Sensitivity (CPVDDS)PAT Model Concordance Rate (PMCR) →

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