(T0930-02-08) Accelerating Development of FDM 3D Printed Medicines with Machine Learning

Purpose: Three-dimensional printing (3DP) is a modern fabrication technology wherein a 3D object is manufactured in a sequential layer-by-layer fashion, based on a computer-aided design (CAD) model [1]. Amongst the different types of 3DP technologies, Fused Deposition Modelling (FDM™) is the most used owing to its affordability, versatility, and compact size [2]. The technology has been used in the pharmaceutical field to produce a variety of drug-loaded products, including printlets, subcutaneous implants, and transdermal microneedles. At present, the empirical process of formulation development hampers advances in pharmaceutical FDM 3DP. An array of parameters within the two-step process influences the printability and performance of the product [3]. Users must therefore undertake a lengthy and costly trial-and-error process of gradually tuning each parameter sequentially before obtaining their desired product. Machine learning (ML) can be employed to optimize this process. ML is an Artificial Intelligence-based technology that enables pattern recognition from large and complex datasets [4]. This study analyzes the performance of three ML techniques for predicting key parameters for fabricating 3D printed drug delivery systems. These ML models are built upon data obtained from published literature and in-house printing experiments (total of 1594 formulations).

Methods: Data mining and collection: Articles published between 1^st January 2013 and 30^th November 2020 were data mined. Additionally, data on in-house formulations, comprising 64 materials and 614 formulations, were included in the dataset. The formulation composition, hot melt extrusion (HME) parameters, FDM parameters, and description of the printed object (including filament mechanical characteristics and overall printability). Feature set creation: Five feature sets were used: material, material name, material type, physical properties, and physical properties per material type (Figure 1). Data analysis and evaluation: Support vector machines (SVM), random forests (RF), and artificial neural networks (ANN) were used in this study to predict the targeted variables: HME temperature, FDM temperature, and printability. ML techniques were developed using Python 3.7, using the Scikit-Learn package. A 75:25 split was used for training and testing. ML techniques used for classification analyses were compared based on their accuracy, Cohen’s kappa, precision, recall and FI. ML techniques used for regression analyses were compared based on their mean absolute error (MAE), and coefficient of determination (R²).

Results: A total of 114 articles were mined, and when combined with in-house formulations, formed a combined dataset comprising 1594 formulations and 260 materials. The combination produced a more diverse dataset, with a broader variety of formulations and range of printing outcomes. Specifically, 72.0% of formulations were printable, which is considerably balanced when considering the general bias reporting of only positive results. Random forest using the “material name” feature set gave the most optimal MAE and R² for predicting HME temperature; 5.54 ºC and 0.91, respectively (Figure 2A). On the other hand, for predicting printing temperature, random forest using the “material with company name” feature set achieved the best MAE and R2 values; 5.99 ºC and 0.88, respectively (Figure 2B). Considering that the printing temperatures ranged between 53 and 240 ºC, the ability to predict the optimal printing temperature within ±5.99 ºC is commendable and would yield notable time saving in formulation and printing development. For classification prediction of printability, random forest using the “physical properties per material type” feature set gave the most optimal performance, with an accuracy of 4%, a Cohen's kappa of 0.66, a F1 score of 0.80, a recall of 80% and a precision of 81% (Figure 3).

Conclusion: Machine learning models built on in-house and published data were able to provide high predictive accuracies for the values of printing and extrusion temperatures, and printability. To support the development of more effective ML models, a structured repository for both positive and negative 3DP data is strongly encouraged. In this way, the development and optimization of FDM formulations may be accelerated, facilitating the advancement of pharmaceutical 3DP.

References: 1. Seoane-Viaño, I., et al., Translating 3D printed pharmaceuticals: from hype to real-world clinical applications. Advanced Drug Delivery Reviews, 2021.
2. Cailleaux, S., et al., Fused Deposition Modeling (FDM), the new asset for the production of tailored medicines. Journal of Controlled Release, 2021. 330: p. 821-841.
3. Melocchi, A., et al., Quality considerations on the pharmaceutical applications of fused deposition modeling 3D printing. International Journal of Pharmaceutics, 2021. 592: p. 119901.
4. Elbadawi, M., et al., Disrupting 3D printing of medicines with machine learning Trends in Pharmacological Sciences, 2021. 42(9): p. 745-757.

Figure 1. Schematic illustrating how materials were classified in the five feature sets.

Figure 2. R2 for (A) extrusion temperatures and (B) printing temperatures using the different ML techniques and feature sets (horizontal axis). RF – random forests, SVM – support vector machines, ANN – artificial neural networks.

Figure 3. Radar plot with metrics results for printability. RF – random forests, ANN – artificial neural networks, SVM – support vector machines.