Track: Formulation and Delivery - Chemical - Formulation - Predictive Modeling
Category: Late Breaking Poster Abstract
Application of Computational Machine Learning Tools to Compare Predictive Performance on Critical Quality Attributes of Spray Dried Lipid-based Formulation
Purpose: The objective of the study was to employ machine learning tools, a broader subset of artificial intelligence (AI) for predicting the critical quality attributes (CQAs) of spray dried self-nanoemulsifying drug delivery system. The main aim of the study was to compare prediction efficiency of different machine learning tools like regression, regularization, reduction, and artificial neural network (ANN). Methods: A Box-Behnken Design (BBD) was employed to optimize the S-SNEDDS. Formulation variables [Drug concentration (DC)] and Aerosil® 200 concentration (AC)] and process variable [Feed rate (FR)] were considered as the independent variables studied at three levels (low, medium, high) with three center points. Percent yield (%) (PY) and loading efficiency (%) (LE) of the S-SNEDDS were considered as CQAs. Minitab® response surface optimizer tool was used to provide prediction optimal solution to achieve maximum PY and LE. Optimized model was verified experimentally by preparing the three replicate batches at the optimized levels of each variables. Regression model: Multiple linear regression (MLR) was employed using Minitab® 18.0 statistical software. Regularization model: Partial least square model (PLS) was applied using Unscrambler® 10.4.1 statistical software. Reduction model: least absolute shrinkage and selection operator (LASSO) regression was implemented by using “glmnet” package in RStudio® statistical software. These three different models: MLR, PLS, and LASSO was employed to quantitatively determine and compare the significant main, interactions, and quadratic effects. An ANN was implemented by using “neuralnet” package in RStudio® statistical software to compare the predictive performance along with the other machine learning tools (MLR, PLS, and LASSO). A one-way ANOVA test (Dunnett’s simultaneous test) was used to compare the prediction performance of these models using Prism 8 GraphPad software. Results: In Figure 1, the mean predicted LE value of LASSO* (91.56%), PLS* (88.85%), and MLR** (92.49%), were statistically different from the experimental verification mean (90.11%). In Figure 1, the mean predicted percent yield (%) value of LASSO* (70.47% predicted mean), PLS**** (71.63% predicted mean), and MLR** (70.74% predicted mean), were statistically different from the experimental verification mean (69.44% measured mean). The predicted mean of ANN model## for LE (91.95% predicted mean) and PY (69.91%) were found to be statistically insignificant showing better coincidence with the mean LE and PY of the verification batches. Overall, MLR, LASSO, PLS, and ANN showed the prediction error less than 5% between the measured and predicted values for both the response variables. Conclusion: ANOVA test (Dunnett test) showed ANN model statistically insignificant between measured and predicted mean for both PY AND LE variables. ANN model showed better predictability as compared to other machine learning models.