Predicting pharmaceutical inkjet printing outcomes using machine learning

Inkjet printing has been extensively explored in recent years to produce personalised medicines due to its low cost and versatility. Pharmaceutical applications have ranged from orodispersible films to complex polydrug implants. However, the multi-factorial nature of the inkjet printing process makes formulation (e.g., composition, surface tension, and viscosity) and printing parameter optimization (e.g., nozzle diameter, peak voltage, and drop spacing) an empirical and time-consuming endeavour. Instead, given the wealth of publicly available data on pharmaceutical inkjet printing, there is potential for a predictive model for inkjet printing outcomes to be developed. In this study, machine learning (ML) models (random forest, multilayer perceptron, and support vector machine) to predict printability and drug dose were developed using a dataset of 687 formulations, consolidated from in-house and literature-mined data on inkjet-printed formulations. The optimized ML models predicted the printability of formulations with an accuracy of 97.22%, and predicted the quality of the prints with an accuracy of 97.14%. This study demonstrates that ML models can feasibly provide predictive insights to inkjet printing outcomes prior to formulation preparation, affording resource- and time-savings.

Accelerating 3D printing of pharmaceutical products using machine learning

Three-dimensional printing (3DP) has seen growing interest within the healthcare industry for its ability to fabricate personalized medicines and medical devices. However, it may be burdened by the lengthy empirical process of formulation development. Active research in pharmaceutical 3DP has led to a wealth of data that machine learning could utilize to provide predictions of formulation outcomes. A balanced dataset is critical for optimal predictive performance of machine learning (ML) models, but data available from published literature often only include positive results. In this study, in-house and literature-mined data on hot melt extrusion (HME) and fused deposition modeling (FDM) 3DP formulations were combined to give a more balanced dataset of 1594 formulations. The optimized ML models predicted the printability and filament mechanical characteristics with an accuracy of 84%, and predicted HME and FDM processing temperatures with a mean absolute error of 5.5 °C and 8.4 °C, respectively. The performance of these ML models was better than previous iterations with a smaller and a more imbalanced dataset, highlighting the importance of providing a structured and heterogeneous dataset for optimal ML performance. The optimized models were integrated in an updated web-application, M3DISEEN, that provides predictions on filament characteristics, printability, HME and FDM processing temperatures, and drug release profiles (https://m3diseen.com/predictionsFDM/). By simulating the workflow of preparing FDM-printed pharmaceutical products, the web-application expedites the otherwise empirical process of formulation development, facilitating higher pharmaceutical 3DP research throughput.

Machine learning predicts 3D printing performance of over 900 drug delivery systems

Three-dimensional printing (3DP) is a transformative technology that is advancing pharmaceutical research by producing personalized drug products. However, advances made via 3DP have been slow due to the lengthy trial-and-error approach in optimization. Artificial intelligence (AI) is a technology that could revolutionize pharmaceutical 3DP through analyzing large datasets. Herein, literature-mined data for developing AI machine learning (ML) models was used to predict key aspects of the 3DP formulation pipeline and in vitro dissolution properties. A total of 968 formulations were mined and assessed from 114 articles. The ML techniques explored were able to learn and provide accuracies as high as 93% for values in the filament hot melt extrusion process. In addition, ML algorithms were able to use data from the composition of the formulations with additional input features to predict the drug release of 3D printed medicines. The best prediction was obtained by an artificial neural network that was able to predict drug release times of a formulation with a mean error of ±24.29 min. In addition, the most important variables were revealed, which could be leveraged in formulation development. Thus, it was concluded that ML proved to be a suitable approach to modelling the 3D printing workflow.