💻 computer science

Artificial Intelligence Framework for Audit-Ready Pharmaceutical Manufacturing Under Explainability, Fairness, and Regulatory Transparency Constraints

This study proposes a comprehensive AI framework for pharmaceutical manufacturing that integrates predictive modeling with explainability, fairness, calibration, drift monitoring, and regulatory traceability to ensure audit-ready, defensible quality-governance decisions.

Original authors: Charles Onyeka Nwamekwe, Nkemakonam Chidiebube Igbokwe, Chukwuma Godfrey Ono

Published 2026-07-22

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Charles Onyeka Nwamekwe, Nkemakonam Chidiebube Igbokwe, Chukwuma Godfrey Ono

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of making medicine as a giant, high-stakes kitchen where chefs (scientists and engineers) are baking billions of pills every day. In this kitchen, the rules are stricter than anywhere else on Earth. If a cookie is slightly burnt, you might just throw it away. But if a pill is even a tiny bit off, it could hurt someone who takes it. Because of this, the kitchen is constantly watched by a team of inspectors who check every single ingredient, every temperature, and every step of the recipe. This is the world of "Good Manufacturing Practice" (GMP), a set of rules that ensures every batch of medicine is safe and perfect.

Now, imagine the kitchen is so big and busy that the human chefs can't watch every single oven at once. So, they bring in a super-smart robot assistant (Artificial Intelligence) to help. This robot can look at thousands of numbers and guess if a batch of pills is going to be bad before it's even finished. But here's the catch: in a normal kitchen, if the robot says "this is bad," you just listen. In the medicine kitchen, the robot can't just shout "Bad!" and walk away. The human chefs and the inspectors need to know why the robot thinks it's bad. They need to see the robot's homework, check if the robot is being fair to all the ovens, and make sure the robot isn't getting confused by old data. If the robot can't explain its thinking clearly, the inspectors won't trust it, and the medicine can't be released. This paper is about building a robot assistant that doesn't just guess, but also explains, stays fair, and keeps a perfect diary for the inspectors.

The Paper: A Robot Chef That Keeps a Perfect Diary

This research paper, written by Charles Onyeka Nwamekwe and his team from Nnamdi Azikiwe University, tackles a big problem: How do we make Artificial Intelligence (AI) safe and trustworthy enough to help make medicine? The authors realized that just having a robot that is good at guessing isn't enough. If the robot is a "black box" (a mystery machine where you can't see how it works), it's useless for making medicine because regulators (the inspectors) won't let it work.

The team designed a new "framework," which is like a rulebook and a blueprint, for building an AI system that is "audit-ready." This means the AI is ready to be inspected at any time, just like a bank account ready for an audit. They didn't just build a model to predict errors; they built a system that checks itself for fairness, explains its own logic, and keeps a detailed log of every decision.

The Ingredients: A Digital Twin of the Factory

To test their idea, the researchers created a massive digital simulation of a pharmaceutical factory. They didn't use real, live patients or real medicine (which would be too risky to mess with), but they used a simulated dataset of 1,200 batches of manufacturing records. Think of this as a video game version of a factory where they could test their robot without ever risking a real pill.

This digital factory tracked everything:

The Recipe: How much moisture was in the mix, how hard the machine pressed the pills, and how fast the machine spun.
The Environment: The temperature and humidity of the room.
The People: Which operator was working, which shift it was (morning, afternoon, or night), and which machine was used.
The Outcome: Whether the batch was perfect or had a defect.

The Magic Trick: Making the Robot Explain Itself

The core of their framework is a special kind of AI that uses a tool called SHAP. Imagine if your robot assistant didn't just say, "This batch is risky," but instead handed you a report card that said: "I think this batch is risky because the moisture was 2% too high, the machine was vibrating a little too much, and the humidity was low."

In their tests, the AI found that granulation moisture (how wet the mix was) was the biggest reason for predicting a problem, followed by compression force (how hard the pills were squished). By showing exactly which factors mattered, the robot stopped being a mystery and became a helpful partner that humans could trust.

The Fairness Check: Is the Robot Biased?

One of the most creative parts of the paper is how they checked for "fairness." Usually, when people talk about AI fairness, they talk about race or gender. But in a factory, fairness looks different. The researchers asked: Does the robot treat every machine, every worker, and every shift equally?

They found that the robot wasn't always fair. For example, it was slightly more likely to make mistakes when judging batches from the "Aba Plant" compared to the "Lagos Plant." It also had a harder time with certain "Product Families" (like Analgesic Tablets) than others. The paper showed that if you don't check for this, the robot might unfairly flag one machine as "broken" just because it's older, or miss a problem on the "Night Shift" because it wasn't trained on enough night data. Their framework catches these biases so they can be fixed before the robot is used for real.

The "Drift" Detector: When the Robot Gets Confused

Factories change over time. Machines get worn out, raw materials change, and seasons shift. This is called "drift." Imagine a robot trained to recognize a fresh apple, but then the apples in the factory start getting slightly different because of a new supplier. The robot might get confused and think a good apple is bad.

The researchers built a "drift monitor" into their system. In their simulation, they watched how the data changed over 24 weeks. They saw that the robot started to get confused around Week 15, when the data patterns shifted significantly. Their system sounded an alarm, telling the humans, "Hey, the world has changed; we need to retrain the robot!" This ensures the AI doesn't slowly become wrong without anyone noticing.

The Final Verdict: The Robot That Keeps a Diary

The most important finding of the paper is that the best AI for making medicine isn't necessarily the one that is the most accurate at guessing. It's the one that is calibrated, explainable, and fair.

They tested several types of AI models. The winner was a model called XGBoost. It didn't just guess right 90% of the time; it also kept a perfect diary. For every single batch it judged, it recorded:

The model version used.
The exact reason for its decision (the "Top Explanation Driver").
Whether it was fair and calibrated.
Who the human reviewer was.
What action was taken (like a "CAPA," which is a fancy term for a "fix-it plan").

In their simulation, 100% of the batches they audited had a complete trail of evidence. The system showed that 93.12% of the time, the documentation was complete enough for an inspector to say, "Yes, this is safe."

What This Means for the Future

The authors are careful to say that this was a simulation. They haven't put this robot in a real, live factory yet. They suggest that the next step is to test it in real-world factories with real machines and real workers.

However, the paper proves a vital point: You cannot just plug a smart computer into a medicine factory and hope for the best. You need a system that is built from the ground up to be transparent. The robot must be able to say, "I think this is bad because of X, Y, and Z," and it must be able to prove that it isn't being unfair to the night shift or the old machines.

By creating this "audit-ready" framework, the researchers have provided a blueprint for how AI can become a trusted partner in saving lives, rather than just a mysterious black box that makes guesses. It's a step toward a future where technology helps us make medicine that is not only effective but also safe, fair, and completely understandable.

Technical Summary: Artificial Intelligence Framework for Audit-Ready Pharmaceutical Manufacturing

Problem Statement
Pharmaceutical manufacturing operates under stringent Good Manufacturing Practice (GMP) regulations where product quality, batch disposition, and patient safety rely on the consistent control of Critical Process Parameters (CPPs) and Critical Quality Attributes (CQAs). While Artificial Intelligence (AI) and Machine Learning (ML) offer potential for real-time quality prediction and predictive maintenance, current deployments face a critical barrier in regulated environments: predictive accuracy alone is insufficient. A model may correctly identify defect risks, but if the prediction lacks explainability, cannot be audited, fails to demonstrate fairness across different production contexts, or lacks calibration, it remains unsuitable for critical GMP decisions. Furthermore, existing literature often treats explainability, fairness, and audit documentation as separate activities rather than an integrated governance structure. There is a significant gap in unified frameworks that simultaneously address prediction, explainability, fairness, calibration, drift monitoring, and human accountability within a single operational lifecycle.

Methodology
The study proposes a conceptual and simulation-ready framework for a "Trustworthy AI" system ( $\mathcal{T}_{AI}$ ) designed specifically for audit-ready pharmaceutical manufacturing. The framework is structured into four integrated layers:

GMP-Compliant Data Layer ( $\mathcal{D}_{GMP}$ ): Utilizes a batch-level dataset of 1,200 records from three manufacturing plants. The data includes CPPs (e.g., granulation moisture, compression force), CQAs (e.g., tablet weight, hardness), environmental variables, and production context (operator, shift, machine, plant, material lot).
Predictive Modelling Layer ( $\mathcal{M}_{AI}$ ): Employs machine learning models (XGBoost, LightGBM, Logistic Regression, etc.) where the optimization objective extends beyond prediction loss ( $L_{pred}$ ) to include fairness loss ( $L_{fair}$ ), calibration loss ( $L_{cal}$ ), explanation-stability loss ( $L_{exp}$ ), and regulatory governance penalties ( $L_{reg}$ ).
Trustworthiness and Governance Layer ( $\mathcal{T}_{TR}$ ): Integrates SHAP-based explainability, Expected Calibration Error (ECE) assessment, manufacturing-specific fairness analysis (evaluating performance gaps across operational groups like shifts and machines), and drift monitoring using statistical divergence measures (Population Stability Index, Jensen–Shannon divergence).
Audit-Ready Decision Support Layer ( $\mathcal{D}_{AUD}$ ): Automatically links every AI prediction to its originating batch record, model version, explanation, calibration status, and Corrective and Preventive Action (CAPA) records to ensure full traceability.

The framework was evaluated using a composite set of metrics: predictive performance (AUC, Macro-F1), trustworthiness (calibration error, fairness gap, drift robustness), and governance effectiveness (audit completeness, CAPA linkage, human override rates).

Key Results

Model Performance: XGBoost demonstrated the strongest overall profile with an AUC of ~0.91 and a composite trustworthiness score of 0.89. While Logistic Regression offered high explanation stability (0.90), it suffered from lower predictive performance (AUC ~0.72).
Calibration and Explainability: Calibration curves indicated that probabilities were generally reliable, though deviations of 0.05–0.08 were observed in mid-risk bins. SHAP analysis identified granulation moisture as the dominant driver of quality risk (mean |SHAP| = 0.17), followed by compression force and relative humidity.
Fairness Analysis: The study revealed that fairness risks are not uniform. The "Product Family" group exhibited the highest fairness gap (0.13), followed by Plant and Operator groups. Machine and Shift groups showed more stable behavior. This confirms that model performance varies significantly across operational contexts, necessitating group-specific monitoring.
Drift and Audit Readiness: Drift detection successfully identified distributional shifts, with Population Stability Index (PSI) crossing alert thresholds in later production windows. The audit dashboard showed 100% model-version traceability and 93.12% audit completeness. In the traceability matrix, all reviewed batches were successfully linked to CAPA records and final dispositions, with audit completeness ranging from 81.98% to 100%.

Significance and Claims
The paper claims to provide the first operational translation of trustworthy AI principles into a unified, GMP-aligned framework for pharmaceutical manufacturing. Its primary significance lies in reframing AI not as a "black-box" automation tool, but as a regulated quality-governance system.

The authors assert that for AI to be defensible in regulatory inspections, every prediction must be treated as a "quality-relevant decision object" that satisfies seven criteria simultaneously: predictive reliability, explanation adequacy, calibration quality, operational fairness, drift stability, CAPA traceability, and human-review accountability. By integrating these elements, the framework positions AI as a system capable of improving audit readiness and regulatory confidence, moving beyond simple accuracy metrics to ensure that AI-supported decisions are transparent, fair, and traceable throughout the product lifecycle.

Limitations
The study acknowledges that the framework was validated using a simulation-ready industrial dataset rather than a live, multi-site GMP deployment. Future work is required to test the system with real-time production data across diverse dosage forms and equipment platforms, and to integrate prospective QA reviewer studies and regulator-facing audit packages.