PAT Adjustment Directional Bias (PADB)

GMI-PH-013Pharmaceutical OperationsPilot-Ready

Definition

Mean signed magnitude of AI-recommended process adjustments, normalized to specification range. Persistent one-directional correction reveals an undetected upstream input bias; an unbiased model trends toward zero.

The Measurement Gap

The Gap: OEE sees nothing. The line is running. Yield sees nothing: no material is wasted. Batch Success Rate sees nothing: batches pass. Process capability sees nothing yet: individual measurements are within range. The problem is that the PAT model is systematically under- or over-adjusting a CPP due to an undetected upstream input bias (e.g., sensor calibration drift). Each recommendation is individually reasonable given the data the model sees. The directional pattern is invisible to every standard metric.

The Decision:

Statistically significant PADB triggers upstream sensor verification and model input audit. PADB normalized above 5% of specification range per adjustment triggers investigation. Above 10% triggers immediate process hold for sensor recalibration. PADB transforms campaign-level model health from an unmeasured assumption into a quantified signal.

Formula

Notation:

PADB = (Σ Adjustment_Magnitude_i × Direction_i) / N_adjustments

Normalize: PADB_normalized = PADB / Specification_Range

Variables:

  • Adjustment_Magnitude_i = the magnitude of each AI-recommended adjustment
  • Direction_i = +1 for upward, -1 for downward
  • N_adjustments = total adjustments over the measurement period
  • Specification_Range = upper limit minus lower limit for the CPP

Worked Example: Over a 34-batch campaign, 356 binder spray rate adjustments. Specification range: 8.0–12.0 mL/min (4.0 mL/min range). Sum of signed adjustments: -28.7 mL/min. PADB = -28.7/356 = -0.0806 mL/min per adjustment. PADB_normalized = -0.0806/4.0 = -2.0% of range per adjustment. t-test: t = -2.58, p = 0.01. Statistically significant. Over a batch with 10 adjustments, cumulative drift = 20% of specification range.

Edge Cases:

  • Report per CPP: bias in one parameter may not exist in others
  • Small campaigns (fewer than 20 batches) may lack sufficient adjustments. Minimum 20 adjustments per CPP for meaningful t-test.
  • PADB = 0 does not mean adjustments are correct: they could alternate in direction and cancel out. Check absolute magnitude as a companion.

Sign convention:

PADB should be near zero. Positive or negative values indicate directional bias.

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher
Reporting Frequency
Continuous; monthly directional-bias review
Category
Autonomous Decision Amplification
← Harvest Timing Attribution Index (HTAI)Adaptive Trial Generalizability Index (ATGI) →

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