(M1530-07-37) Optimization of 3D Printed Tablets Using Artificial Intelligence and Machine Learning

Purpose: Three-dimensional (3D) printing is one of the most promising technologies for fabricating personalized dosage forms for patients [1]. Artificial Intelligence (AI) consists of neural networks which are often used in pharmaceutical research, particularly Artificial neural networks (ANN) which are used in the optimization [2]. The process of drug development to commercialization of the dosage form takes at least a decade or more, clinical trials take approximately 6-7 years. The development of a novel dosage form cost $2.6 billion for the pharmaceutical companies [3]. The integration of artificial intelligence with 3D printing can help researchers achieve high-quality drug products and eventually overcome issues such as time, cost, and need for resources. The current study was performed to develop optimized fused deposition modeling (FDM) based formulations using batch and continuous printers.

Methods: A factorial design was applied to print formulations with different parameters using commercial polylactic acid (PLA) filament. Each printer (batch and continuous) printed 81 formulations, yielding physical dimensions and images. Image segmentation was performed to calculate the porosity which was later utilized for predicting parameters (Figure 1 (A)). The captured images were converted to grayscale using Python to find the threshold value which separated the pixel values of the background from the boundary of the tablet. The threshold values of 0.45 and 0.35 were determined for batch and continuous printers respectively. Random forest was applied because of its excellent predictive performance and low overfitting issue. It consisted of hundreds of decision trees which were built by both random extraction of observations and a random extraction of features from the dataset. The random forest result is shown in Figure 1 (B), where the x-axis is the feature importance, and the y-axis is the list of processing parameters. From this figure, the processing parameters are ranked by feature importance value from high to low, presenting the competition of flow %, infill density, print speed, and print temperature in affecting the sample surface quality. It is observed that flow % occupies a much more dominant position over the other three processing parameters in predicting the segregation energy. The operating principle of adaptive design (Figure 2 (C)): the initial dataset is formed with a porosity target calculated by image segmentation. Gaussian regression models are trained using the initial dataset. The trained model will be applied to the search space of processing parameters to generate uncertainties. Hence selector (Efficient Global Optimization (EGO)) will work on selecting the next set of processing parameters for the experiment. If the sample demonstrates the desired porosity, this set of processing parameters is output as the design achievement, and the loop ends. Otherwise, the sample will be added to the current dataset and utilized in the next loop. The loop will keep running until the targeted porosity is selected. In the adaptive design, a trade-off between exploration (global search) and exploitation (local optimization) is balanced in suggesting the next experiment.

Results: A new set of tablets were printed, which were based on the predictions made by the algorithm. The algorithm predicted novel formulations, which included parameters not found in the dataset. Table 1 (A) (Batch) and (B) (Continuous) depicts the predictions with 0% defects explored by machine learning and physical dimensions of 3D printed tablets. The algorithm predicted that infill density had a significant impact on the tablets' weight, while flow rate affected both the tablets' quality and their weight. In Table 1 (A), F3 and F5 had high weight among all the formulations because of high infill density and flow rate. F2 had the lowest weight among formulations even though F4 had an infill density lower than F2 but F4 had a higher flow rate. Table 1 (B) depicts similar observations such as F1 and F8, F2 and F4, F5 and F6, and F3 and F7 had the same infill density but because of flow rate, the weight of the tablet was different. Figure 2 (A) depicts the images printed using a batch printer.  The images show that a flow rate of 110-120% for batch printer showed high-quality tablets and there were no observed defects on the surface of the tablets. Figure 2 (B) depicts tablets printed using continuous printer, the tablets showed no observed defects on the surface of tablets with a flow rate ranging from 65- 120 %.

Conclusion: The algorithm was successful in predicting printlets with an absence of defects. The current research showed that infill density and flow rate are crucial parameters that define the weight of the tablet as well as the overall quality of the tablet. The algorithm can be implemented for drug-loaded formulations as well to minimize the time and resources required for optimizing printing parameters. This novel method of fabricating 3D-printed tablets holds significant promise for future research, potentially leading to groundbreaking advancements in the field of personalized medicine.

References: 1. Mathew, E.; Pitzanti, G.; Larrañeta, E.; Lamprou, D.A. 3D Printing of Pharmaceuticals and Drug Delivery Devices. Pharmaceutics 2020, 12, 266.
2. Bourquin, J.; Schmidli, H.; Van Hoogevest, P.; Leuenberger, H. Basic Concepts of Artificial Neural Networks (ANN) Modeling in the Application to Pharmaceutical Development. Pharmaceutical Development and Technology 1997, 2, 95–109,
3. Avorn J. The $2.6 billion pill--methodologic and policy considerations. N Engl J Med. 2015 May 14;372(20):1877-9

Table 1: Predicted Parameters and Physical Dimensions of 3D Printed Tablets [A] Batch [B] Continuous

Figure 1 – [A] Example of Porosity Calculation Process. (a) Original photograph of tablet; (b) Segmentation of tablet (yellow) from background (purple); (c) Shaded regions of tablet (blue) and voids (red) for a grayscale threshold of 0.45, yielding a porosity of 0.07, [B] Feature importance for each processing parameters and [C] Graphical Abstract

Figure 2 - Surface of 3D Printed Tablet from Experiments Guided by Predictions [A] Batch Printer [B] Continuous Printer [C] Adaptive Design Loop