Related Experiment Videos
Comparison of artificial neural networks (ANN) with classical modelling techniques using different experimental
J Bourquin1, H Schmidli, P van Hoogevest
1Pharmaceutical and Analytical development, Novartis Pharma AG, CH-4002 Basel, Switzerland. jacques.bourquin@pharma.novartis.com
Summary
Artificial Neural Networks (ANN) modeling offers superior robustness for analyzing tabletting data, especially from randomized trials, compared to Response Surface Methodology (RSM). ANN models are less sensitive to experimental design, making them ideal for historical or less organized data.
Area of Science:
- Pharmaceutical Sciences
- Computational Chemistry
- Data Science
Background:
- Traditional Response Surface Methodology (RSM) is widely used for optimizing pharmaceutical processes.
- Analyzing experimental data from complex studies, like tabletting, requires robust modeling techniques.
- The organization of experimental design can significantly impact model reliability.
Purpose of the Study:
- To quantitatively compare Artificial Neural Networks (ANN) with RSM for analyzing tabletting experimental data.
- To evaluate the data fitting and model robustness of ANN and RSM under different experimental designs (statistical vs. randomized).
- To determine the suitability of ANN for analyzing historical or poorly organized trial data.
Main Methods:
- Artificial Neural Networks (ANN) methodology was employed for data analysis.
- Response Surface Methodology (RSM) was used as a classical benchmark for comparison.
- Two experimental designs were utilized: a statistical, organized design and a randomized design.
- Goodness of fit was assessed using the R² coefficient; model robustness was evaluated using an independent validation dataset's R².
Main Results:
- Both ANN and RSM yielded comparable results when applied to data from a statistical experimental plan.
- ANN methodology demonstrated superior model robustness compared to RSM when data originated from a randomized experimental plan.
- ANN models showed less sensitivity to the organizational level of the trial design.
- Tablet properties were primarily influenced by compression dwell time, silica aerogel concentration, and magnesium stearate concentration.
Conclusions:
- Artificial Neural Networks (ANN) methodology is a more robust and adaptable tool for analyzing tabletting data, particularly when dealing with randomized or historical experimental designs.
- ANN's reduced sensitivity to experimental design organization makes it a valuable alternative to RSM for complex pharmaceutical data analysis.
- Dwell time, silica aerogel, and magnesium stearate are critical factors influencing tablet properties.