[Sparse Denoising Autoencoder Application in Identification of Counterfeit Pharmaceutical].

Area of Science:

• Analytical Chemistry
• Machine Learning
• Pharmaceutical Analysis

Background:

• Near-infrared (NIR) spectroscopy offers a fast, non-destructive method for pharmaceutical discrimination.
• Autoencoder networks, a type of deep learning model, demonstrate superior modeling capabilities compared to traditional algorithms.
• Effective feature extraction and dimensionality reduction are crucial for accurate pharmaceutical analysis.

Purpose of the Study:

• To develop and evaluate a Sparse Denoising Autoencoder (SDAE) model for identifying counterfeit pharmaceuticals.
• To investigate the impact of pre-whitening transformation on the performance of the SDAE model for near-infrared spectroscopy data.
• To compare the classification accuracy and generalization performance of the SDAE model against BP neural networks and Support Vector Machine (SVM) algorithms.

Main Methods:

• Near-infrared spectroscopy data of erythromycin ethylsuccinate was preprocessed and subjected to pre-whitening transformation.
• A Sparse Denoising Autoencoder (SDAE) with two hidden layers was constructed and trained using unsupervised greedy layer-wise pre-training followed by supervised fine-tuning.
• The SDAE model's performance was evaluated and compared with BP neural networks and SVM algorithms based on classification accuracy and mean absolute difference (MAD).

Main Results:

• Pre-whitening transformation effectively reduced feature correlation and improved the classification accuracy of the SDAE model.
• The SDAE algorithm achieved higher classification accuracy than both BP neural networks and SVM algorithms, especially with larger training datasets.
• The SDAE model demonstrated superior generalization performance, indicated by a lower mean absolute difference in classification accuracy compared to SVM and BP neural networks.

Conclusions:

• The proposed SDAE model, combined with pre-whitening, is an effective tool for the accurate discrimination of counterfeit pharmaceuticals using NIR spectroscopy.
• SDAE offers improved classification accuracy and generalization capabilities over traditional machine learning algorithms for this application.
• This approach provides a robust method for ensuring pharmaceutical quality and authenticity.