Predictive Intervention Value Index (PIVI)

GMI-PH-025Pharmaceutical OperationsIntegration-Required

Definition

Proportion of AI-triggered predictive interventions where matched historical trajectory analysis confirms the batch was genuinely heading toward failure. Separates interventions that saved a batch from interventions that took credit for a batch never at risk.

The Measurement Gap

A bioreactor predictive quality model triggers 47 interventions in a quarter. Every flagged batch passes. The model gets credit for saving them all. But nobody checks whether those batches would have passed without intervention. A model that flags every batch and triggers interventions on all of them would show 100% of flagged batches passing, and by that standard looks perfect.

Batch success rate only tells you the batch passed or failed after the fact. It does not distinguish between a batch that passed because of an intervention and one that passed despite it. In pharma, an unnecessary intervention during a biologic batch is a potential source of uncontrolled variability introduced into a GMP-regulated process.

Decision enabled:

When PIVI drops below 0.6, the model is triggering more disruption than help. Raise the confidence threshold at which the model triggers interventions, or retrain the model. When PIVI is high, the interventions are genuine corrections and batch success improvements are real.

Formula

PIVI = interventions_with_validated_value / total_predictive_interventions

Where interventions_with_validated_value counts interventions where matched historical trajectory analysis confirms the batch was heading toward failure. Total_predictive_interventions counts all instances where the predictive model triggered a manual process adjustment.

Worked Example:

A monoclonal antibody site runs 186 batches in Q3. The predictive model triggers interventions on 34. For each, the analytics team identifies 5–8 historical batches with similar dissolved oxygen, pH, and viable cell density trajectories at the trigger point.

19 interventions: historical comparators had 60–90% failure rates. Intervention plausibly prevented failure (validated value).
9 interventions: historical comparators had below 15% failure rates. Batches were likely passing without intervention (unnecessary disruption).
6 interventions: insufficient historical comparators. Excluded.

PIVI = 19 / (19 + 9) = 19 / 28 = 0.679

Above the 0.6 threshold but the 9 unnecessary interventions each introduced uncontrolled variability into a GMP process.

Sign Convention:

Higher is better. Above 0.6 indicates net value. Below 0.3 means the model triggers more disruption than help.

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher
Reporting Frequency
Per intervention; quarterly value review
Category
Predictive Manufacturing Quality
← Inspection Performance Throughput Decay (IPTD)Quality Prediction Confidence Index (QPCI) →

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