Convolutional neural network-based evaluation of chemical maps obtained by fast Raman imaging for prediction of tablet dissolution profiles.

In this work, the capabilities of a state-of-the-art fast Raman imaging apparatus are exploited to gain information about the concentration and particle size of hydroxypropyl methylcellulose (HPMC) in sustained release tablets. The extracted information is utilized to predict the in vitro dissolution profile of the tablets. For the first time, convolutional neural networks (CNNs) are used for the processing of the chemical images of HPMC distribution and to directly predict the dissolution profile based on the image. This new method is compared to wavelet analysis, which gives a quantification of the texture of HPMC distribution, carrying information regarding both concentration and particle size. A total of 112 training and 32 validation tablets were used, when a CNN was used to characterize the particle size of HPMC, the dissolution profile of the validation tablets was predicted with an average f2 similarity value of 62.95. Direct prediction based on the image had an f2 value of 54.2, this demonstrates that the CNN is capable of recognizing the patterns in the data on its own. The presented methods can facilitate a better understanding of the manufacturing processes, as detailed information becomes available with fast measurements.

Raman mapping-based non-destructive dissolution prediction of sustained-release tablets.

In this paper, the applicability of Raman chemical imaging for the non-destructive prediction of the in vitro dissolution profile of sustained-release tablets is demonstrated for the first time. Raman chemical maps contain a plethora of information about the spatial distribution and the particle size of the components, compression force and even polymorphism. With proper data analysis techniques, this can be converted into simple numerical information which can be used as input in a machine learning model. In our work, sustained-release tablets using hydroxypropyl methylcellulose (HPMC) as matrix polymer are prepared, the concentration and particle size of this component varied between samples. Chemical maps of HPMC are converted into histograms with two different methods, an approach based on discretizing concentration values and a wavelet analysis technique. These histograms are then subjected to Principal Component Analysis, the score value of the first two principal components was found to represent HPMC content and particle size. These values are used as input in Artificial Neural Networks which are trained to predict the dissolution profile of the tablets. As a result, accurate predictions were obtained for the test tablets (the average f2 similarity value is higher than 59 with both methods). The presented methodology lays the foundations of the analysis of far more extensive datasets acquired with the emerging fast Raman imaging technology.