(T1430-08-44) Application of Machine Learning Tools for Predicting Impact of Punch Cup Depth on Tablet Capping

Purpose: The aim of the study was to study the ability of newly developed dimensionless ‘capping index (CI)’ parameter to evaluate capping impact of punch cup depth on the model acetaminophen tablets. The generated data set was used to develop a predictive capping index-based model using machine learning tools and multivariate analysis.

Methods: Different punch cup depths [flat face (FF), flat face beveled edge (FFBE), flat face radius edge (FFRE), standard concave (SCC), shallow concave (SHCC), compound concave (CCC), deep concave (DCC), and extra deep concave (XDCC)] were used for manufacturing acetaminophen tablets. This formulation was composed of acetaminophen, microcrystalline cellulose, copovidone, silica, and magnesium stearate (50:43:5:1:1). Tablet were manufactured at 50, 100, 200, 250, and 300 MPa compression pressure at 20 RPM of different cup depth tools using Natoli NP RD 30 tablet press. Tablet capping were observed during the tablet breaking study. Total 10 tablets were tested during this study. An axial and a radial tablet fracture study were performed on the Gamlen D1000 for measuring of breaking force and Young’s moduli. CI is a ratio of Mean Isotropic Index (MII) and Mean Anisotropic Index (MAI). MII is the ratio of an axial and a radial breaking force. MAI is the ratio of an axial and a radial Young’s moduli. A quantitative partial least square model (PLS) was computed [Unscrambler^® 10.4.1, Trondheim, Norway] to understand the effect of cup depths and compression parameters on the CI. A classification tool in the machine learning algorithm such as Decision Tree was used to model the impact of the cup depth and compression parameters on the capping observation as ‘YES or NO’ response. RStudio^® statistical software’s (Boston, MA) “rpart” package (Version 4.0-2) was implemented to develop a Decision Tree. In-house script was developed for Decision Tree modeling of the capping observation data.

Results: A PLS model (Figure 1) exhibited a positive impact of the punch cup designs such as DCC and XDCC on the tablet CI. A compression pressure and an ejection force displayed a respective positive and negative impacts on the tablet CI. A Decision Tree (Figure 2) confirmed CI as a prominent variable for tablet capping. In this model, 94% events in the Cluster 1 led to intact tablets with CI < 0.41, while 82% capping occurrence was observed in the Cluster 2 with CI > 0.41.

Conclusion: A study revealed that tablet capping is a direct function of an increment in the punch cup depth and cup volume of punches in addition to the intresnic material properties of the formulation blend. A developed Decision Tree could serve as a ‘guiding tool’ for proper selection of the cup design to avoid potential tablet problems.

Figure 1: Partial least square (PLS) regression analysis of a Capping Index [ Flat Face (FF), Flat Face Beveled edge (FFBE), Flat Face Radius Edge (FFRE), Standard Concave (SCC), Shallow Concave (SHCC), Compound Concave (CCC), Deep Concave (DCC), and Extra- Deep Concave (XDCC)].


Figure 2: Decision Tree model of capping observation (YES/NO).