Deep Learning-Driven Preprocessing of FTIR Spectra via 1D-CNN for Antibiotic Monitoring

Abstract

Fourier-transform infrared spectroscopy (FTIR) has emerged as a rapid, non-destructive tool for the characterization and quality monitoring of pharmaceuticals. An approach that integrates deep learning for spectral data preprocessing and unsupervised learning for the analysis of beta-lactam antibiotics, specifically amoxicillin and ampicillin, in both intact and degraded forms, is presented in this study. In the first step, raw FTIR spectra underwent an automated preprocessing workflow using 1D convolutional neural networks (1D-CNN: 32, 64, 128, and 64 filters, kernel sizes 7 and 5) to optimize and standardize data quality and reduce spectral interference. The K-means clustering algorithm, applied to the automated preprocessed spectra, identified similarity patterns among samples and grouped them by antibiotic type and degradation state, without prior knowledge of their classes. The Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) were employed to compare the generated clusters against the reference classification of the samples, evaluating the quality of the resulting clusters, and the results (ARI = 0.8421 and NMI = 0.8835) showed high agreement between the CNN-generated clusters and the expected classes. The findings of this study demonstrate that the automated data preprocessing combination preserved relevant spectral information, enabling efficient discrimination between intact and degraded samples of the studied antibiotics. Furthermore, the integration of FTIR spectroscopy, chemometrics, and AI leads to a promising strategy for pharmaceutical quality analysis.

Speaker

Vinicius P. Anjos
Instituto de Pesquisas Energéticas e Nucleares, Universidade de São Paulo.
Brazil

Discussion

Ask question