Application of support vector machine learning for orodispersible films disintegration time prediction

Abstract

1. INTRODUCTION Orodispersible films (ODF) have emerged as innovative dosage forms that provide wide variety of advantages for patients and manufacturers over conventional dosage forms. The prominent characteristic of ODFs is fast disintegration followed by good patients acceptability [1]. Therefore, relevant disintegration time (DT) is usually considered as ODF critical quality attribute. Extensive research on ODFs is generating a lot of data, but lack of standardization is the main obstacle that limits their comparative evaluation. The following work aims to explore literature data on ODFs characteristics using the predictive data-classification algorithm Support vector machine (SVM) and assess its applicability in pharmaceutical development based on the set of experimentally obtained data. 2. MATERIALS AND METHODS 2.1. Materials Hydroxypropyl cellulose (Klucel GF, Ashland, USA), ethanol (≥99.8%, Honeywell, Charlotte, NC, USA) and glycerol, 85% (w/w) (Ph. Eur.) were used for preparation of printing and casting dispersion. 2.2. Data pre-processing Comprehensive data exploration has been conducted in the PubMed database using most common synonyms for ODFs with fifteen synonyms in singular and plural. Built database had following attributes: manufacturing approach, polymer selection, polymer molecular weight (KDa), polymer load (%), mechanical properties (tensile strength (MPa), Young's modulus (MPa), elongation at break (%)), disintegration method and disintegration time (DT) (s). 2.3. ODF preparation and characterisation Polymer dispersions for solvent casting and semi-solid extrusion 3D printing were prepared by dispersing HPC in ethanol:glicerol solution followed by continuous stirring on the magnetic stirrer. Prepared dispersions were: (i) casted on a unit-dose plexiglas plates, or (ii) printed using Ultimaker 2+ (Ultimaker, , Netherlands). ODFs were characterized in terms of mechanical properties using Z-LX Table-Top Testing Machine (Shimadzu, Japan) and DT using adapted compendial tester (Erweka ZT52, Germany) with a weight. 3. RESULTS AND DISCUSSION 3.1. Data pre-processing 274 papers (without reviews) were identified via search, of which 112 were included in the database. Nominal data from literature was transformed into numerical, using coding operator so that each nominal data had corresponding numerical value. Critical attributes for films fast disintegration were derived. 18 polymers were included as categorical data and were further differentiated on the basis of molecular weight. Values for most commonly evaluated mechanical properties were included as numerical data. Different DT methods were classified in seven classes (Table 1), while the manufacturing methods were classified in five classes. RapidMiner Studio 9.10 (RapidMiner, Dortmund, Germany) was used to transform data and employ SMV algorithm. 3.2. SVM model prediction Attributes with the highest weight were polymer load and DT method employed (Figure 1). The polymer type and characteristic did have conclusive effects on DT as their weight varied during data mining. This can be attributed to inconclusive data provided in papers and lot of missing values for those attributes. Mechanical properties had low weight, which can be explained with the broad value range for those attributes. Different research groups had different approach to disintegration testing, which lowered model precision as it was reported that SVM does not have high accuracy when data is imbalanced [3]. Relative error value was 20%, which can be considered as high, but, having in mind great diversity in presented data and methodology, obtained value is still acceptable for the pilot study. 3.3. Experimental validation HPC-based films prepared by 3D printing had tensile strength, elongation at break and Young’s modulus of 3.5 MPa, 137% and 5 MPa, respectively. Average DT was 69 s. For casted films, relevant values were 3.4 MPa, 105% and 3 MPa, and DT was 27 s. Experimentally obtained results were entered into model simulator (Figure 2) to simulate situation reflecting the experimental set up in which HPC-based films were prepared by 3D printing and solvent casting, and relevant attribute values obtained by samples characterization. In the case were manufacturing method was set to be 3D printing (coded as 1) predicted DT value was close to experimentally obtained value, i.e. 71.7 and 69 s, respectively. When solvent casting method was considered, predicted DT value was remarkedly higher than the experimentally obtained one, indicating bad predictability. It might be assumed that good predictability obtained in the case of 3D printed films is associated with lower data variability due to more simple sample composition and robust preparation method. In the case of casted films, data was much more complex due to a higher number of research papers and approaches to characterisation. 4. CONCLUSION The obtained results indicate that SVM algorithm can be employed to predict ODF DT value based on the dataset created using literature data. However, in order to obtain meaningful predictions, larger dataset, with fewer inconsistences and less missing values would be advantageous.