Machine Learning-Enabled NIR Spectroscopy. Part 3: Hyperparameter by Design (HyD) Based ANN-MLP Optimization, Model...

Area of Science:

• Data Science
• Machine Learning
• Pharmaceutical Science

Background:

• Data-driven models face failures due to variations, leading to model drift and inaccurate predictions.
• Monitoring and mitigating model drift are critical for maintaining predictive accuracy in data-driven applications.
• Regulatory bodies like the USFDA and ICH emphasize risk-based approaches for managing pharmaceutical variations.

Purpose of the Study:

• To investigate hyperparameter optimization for Artificial Neural Network Multilayer Perceptron (ANN-MLP) models.
• To implement a robust methodology combining Design of Experiments (DoE), drift analysis, and Statistical Process Control (SPC).
• To ensure model reliability and accurate predictions by managing data variations and model lifecycle.

Main Methods:

• Employed Design of Experiments (DoE) for pre-screening and optimization on internal validation data to define design and control spaces.
• Utilized regression performance metrics to select optimal hyperparameters, optimizing modeling time and storage.
• Performed target drift analysis on external validation data and trend analysis on mean absolute error using SPC.

Main Results:

• Identified optimal hyperparameters through DoE, balancing performance with computational efficiency.
• Observed drift in external data but confirmed it remained within the validated range of internal data.
• Established out-of-specification and process control limits, demonstrating the model's reliability and performance.

Conclusions:

• The integrated approach of DoE, drift analysis, and SPC enables robust hyperparameter optimization and effective model lifecycle management.
• This methodology ensures accurate and dependable predictions by validating model performance against internal and external data.
• The study provides valuable insights for maintaining model integrity and reliability in real-world applications, aligning with regulatory expectations.