(M1130-02-12) Machine Learning Predicts Pharmaceutical Inkjet Printing Outcomes

Purpose: Inkjet printing has garnered considerable attention within the drug delivery space for its versatility in producing personalised medicines and unique dosage forms. However, in pharmaceutical inkjet printing research, a significant amount of time and resources is invested into optimization of ink characteristics and printing parameters. Due to the large number of variables, ink formulation development is an empirical process that is time-, material-, and cost-consuming. Conventional guidance on determining the inkjet printability of an ink is based on the Ohnesorge number, which is a dimensionless number that describes the tendency of a droplet to stay intact. An Ohnesorge number ranging between 0.1 and 1, which is equivalent to a Z value (where Z is the reciprocal of the Ohnesorge number) of 1–10, is often considered to be printable. However, there have been numerous exceptions, with inks with Z values above 10 found to be printable. Therefore, a predictive tool to better determine the printability of inks a priori would allow faster ink formulation development. This study investigates the use of machine learning (ML) to predict the printability of pharmaceutical inkjet printing inks, based on information collated from published articles and in-house experiments (total of 687 formulations). In doing so, the study aims to develop an in-silico tool for better prediction of inkjet printability to help accelerate pharmaceutical inkjet printing formulation.

Methods: Data collection: Articles using the terms “inkjet printing” or “ink jet printing” or “ink-jet printing”, and “drug” or “drug device”, published between May 2000 and February 2022 were consolidated. Additionally, information from several in-house formulations were included. Data was collected on (1) formulation composition and properties, (2) individual material properties, (3) process-related parameters, and (4) target variables. Feature set generation: Five feature sets were used: material, material name, material type, physical properties, and physical properties per material type. ML model development and performance analysis: Three ML techniques were used in this study: multilayer perceptron (MLP), support vector machines (SVM), and random forests (RF). Every possible permutation of MLT, feature set, and relevant additional input parameters were evaluated for each target variable over 50 random seed values. A 75:25 split was used for training and testing the MLTs. The performance of ML models developed from each permutation was evaluated based on their Cohen’s kappa, precision, recall, and F1. The permutation that performed the best for each target variable was subsequently optimized by hyperparameter tuning (grid search with 5-fold cross-validation).

Results: A total of 687 formulations were extracted from 75 articles and 2 in-house projects, of which 636 were printable (92.6% of all formulations) (Figure 1). Amongst the printable formulations, 30 produced satellite droplets (4.72% of printable formulations). This positively skewed dataset highlights the tendency for researcher to publish only positive results. Additionally, only 68 printable formulations had Z values > 10. Rather than supporting the guidance on Z value, this clustering of Z values between 1 – 10 may instead indicate a routine omission of formulations with Z values > 10 from printing. The best performing ML model for predicting whether inkjet printing formulations were printable was based on a RF algorithm, using the “grouped by material and company” feature set, coupled with information on the nozzle diameter, nozzle, and printer (Figure 2A). This configuration gave a dataset comprising 429 formulations (62.4% of all formulations), of which 9.32% were negative outcomes. The optimized model had an accuracy of 97.22% and a Cohen's kappa coefficient of 0.854. The best performing model for predicting whether jettable formulations produce satellite droplets was based on a MLP algorithm, using the “grouped by material type” feature set, coupled with information on nozzle diameter and printer (Figure 2B). This configuration resulted in a residual data size of 412 formulations (64.8% of printable formulations), 1.70% of which were negative outcomes (satellite droplets). Comparing the conventional guidance that only inks with Z values < 10 printable, the model developed in this study performed significantly better. Based on the same dataset used for training and testing the ML model for predicting printability, conventional guidance produced a false positive rate of 77.42%, a false negative rate of 28.16%, and only an accuracy of 64.39% (vs 97.22% using the ML model) (Table 1). This highlights the multifactorial dependence of inkjet printing outcomes, and the importance of considering numerous material- and process-related parameters beyond a formulation's Z value in making such predictions.

Conclusion: ML models built on published and in-house inkjet printing data provided high predictive accuracies on inkjet printing outcomes that outperformed conventional guidance. With more well-balanced dataset, achieved through open reporting of negative results, models with better performance and generalizability can be developed. This will in turn accelerate inkjet printing formulation development.

Acknowledgements: L.R.P. acknowledges the predoctoral fellowship provided by the Ministerio de Universidades (Formación de Profesorado Universitario (FPU 2020). I.S.V. acknowledges Consellería de Cultura, Educación e Universidade for her Postdoctoral Fellowship (Xunta de Galicia, Spain; ED481B-2021-019).

Figure 1. Sankey diagram showing distribution of printability

Figure 2. Radar plot with metrics results of models predicting (A) printability (yes or no), and (B) printability (good or satellite).

Table 1. Confusion matrix based on conventional guidance.