Materials

The active pharmaceutical ingredient, clopidogrel hydrogen sulphate was obtained from MSN Laboratories Ltd (Telangana, India). The excipients used for fluidized hot melt granulation (FHMG) consisted of Mannitol 35 (Roquette Frères, Lestrem, France) and cellulose, microcrystalline type 103D + (Mingtai Chemical Co., Ltd., Taoyuan City, Taiwan) as fillers, Macrogol 8000 (Dow Chemical Company, Hahnville, LA, USA) as plasticizer, and low-substituted hydroxypropyl-cellulose L-HPC, LH-11 (Shin-Etsu Chemical Co., Ltd., Tokyo, Japan) as a binder. Kolliwax HCO (BASF SE, Ludwigshafen, Germany), a hydrogenated castor oil, served as a lubricant. All materials were of Ph. Eur. grade. Potassium chloride, potassium hydroxide, and hydrochloric acid (Merck GmbH, Darmstadt, Germany) were of analytical grade and used to prepare dissolution media.

Pharmaceutical Composition and Manufacturing Technology

The pharmaceutical composition of Clopidogrel 75 mg tablets is given in Table I.

Table I Qualitative and Quantitative Composition of Clopidogrel 75 mg Tablets

The manufacturing technology comprised of the following steps:

1.

Sieving – Mannitol 35 and Macrogol 8000 were passed through a sieve insert with 1.0 mm pore size using a Frewitt-Coniwitt Lab type rotary sieve (Frewitt Ltd., Fribourg, Switzerland).

2.

FHMG – the sieved mixture alongside the active pharmaceutical ingredient, microcrystalline cellulose, and L-HPC were loaded into a Bosch Solidlab 1 fluid–bed granulator and dryer (Bosch GmbH, Schopfheim, Germany) and granulated according to the setpoints defined in the experimental design (detailed in Sect. 2.3).

3.

Sieving II – the cooled granules were passed through a sieve insert with 1.0 mm pore size.

4.

Lubrication – the resulted sieved granules were lubricated with Kolliwax HCO using an Erweka AR402 double-cone blender (Erweka GmbH, Langen, Germany).

5.

Tablet compression – the obtained lubricated granules were compressed to tablets using round, biconcave punches with a diameter of 8.0 mm using a RIVA Piccola D8 rotary tablet press (RIVA S.A., Ciudadela, Buenos Aires, Argentina).

The manufacturing steps and in-process control steps are summarized in Fig. 1.

Fig. 1

Manufacturing process steps and control points for Clopidogrel 75 mg tablets

Product Optimization Strategy

The product and process optimization of Clopidogrel 75 mg tablets was based on the pharmaceutical particulars of the active substance and incorporated excipients. To encompass the differences in dissolution behavior of clopidogrel based on the quantity of plasticizer and according to variations in the manufacturing technology a calibration set of 15 laboratory scale experiments was conducted. The span of granulation temperature was selected considering the melting range of Macrogol 8000, namely 60–63°C and granulation time was set between 5–30 min. Macrogol content varied between 8–18% of the target formulation. The differences in Macrogol content were corrected by modifying the quantity of microcrystalline cellulose. The full – factorial experimental design with three center points is presented in Table II. The resulting granules were divided into two equal parts and lubricated for 2 and 15 min with Kolliwax HCO, respectively. A short lubrication time of 2 min was chosen to reflect typical practices in the pharmaceutical industry, while a 15-min duration was selected to represent an extreme condition under which over-lubrication may occur. During formulation development, it is crucial to consider factors that may impact the downstream processability of granules. Excessive lubrication of powder blends can lead to suboptimal tablet compression, extended disintegration times, and impaired API release from the final dosage form. The obtained lubricated granules were compressed to tablets at three different compression forces of 10, 17.5 and 25 kN, maintaining the turret/feeder rotation speed ratio at 1 for all tests, thus applying six post-granulation conditions to each of the 15 granulation batches. This yielded a total of 90 unique process settings, from which three tablets per setting were selected for dissolution testing. The range of compression forces was established to cover the values which are sufficient to achieve acceptable tablet tensile strength, extending to levels at which a plateau in this in-process parameter was observed. Additionally, the upper limit was constrained by the maximum tip pressure tolerated by the compression tooling. Accordingly, a force range of 10–25 kN was selected, with 17.5 kN representing the arithmetic mean.

Table II Worksheet of the Conducted Experimental DesignIn-Process Control Testing of Granules and Tablets

In-process control of granules included the determination of flow-out time, bulk density, and angle of repose using a PharmaTest PTG-S4 powder testing system (PharmaTest Apparatebau AG, Hainburg, Germany). The particle size distribution of granules was determined using the Vibratory Sieve Shaker Analysette 3 PRO (Fritsch GmbH, Weimar, Germany) using test sieves of 80–200–400800–1000 µm, with a vibration frequency of 1 mm for 10 min. The loss on drying of the lubricated granules was measured with a Mettler Toledo HR-73-type halogen moisture analyzer (Mettler Toledo, Columbus, OH, USA) using a standard drying method of 70°C for 20 min. Tablet mass, resistance to crushing, diameter, and height were measured using a PharmaTest WHT3ME automated tablet testing system (PharmaTest Apparatebau AG, Hainburg, Germany). The friability of tablets was determined with a PharmaTest PTF 10 E friability tester (PharmaTest Apparatebau AG, Hainburg, Germany). The disintegration time was determined using a PTZ Auto 1 Single Position Semi-Automated Disintegration Testing Instrument (Pharma Test Apparatebau AG, Hainburg, Germany).

Fourier Transformation Near-Infrared Spectroscopy

Near-infrared spectra of the tablets and granules were acquired using a Bruker MPA II FT-NIR Spectrometer (Bruker, Billerica, MA, USA). The reflectance spectra of the granules have been collected using an integrating sphere unit, while the transmission and reflectance spectra of the tablets was obtained with the tablet measurement unit, with both sides being measured. Spectral data were obtained from both sides of each tablet. Reflectance spectra were recorded using 32 scans and data acquisition was done in the 3952–11536 cm–1 spectral range, with a resolution of 16 cm–1. Transmittance spectra were collected in the spectral range of 5808–11520 cm–1 employing 64 scans with a resolution of 32 cm–1. The final spectra were evaluated as the average of the consecutive scans. Background reference was recorded before the measurements.

In Vitro Dissolution Testing

Dissolution testing was carried out using an Erweka DT800 automated dissolution tester system (Erweka GmbH, Langen, Germany). In vitro dissolution of Clopidogrel followed the United States Pharmacopoeia monograph for Clopidogrel Tablets (45). Dissolution profiles were recorded in 1000 mL KCl/HCl buffer dissolution media at pH = 2.0 ± 0.05 at 37°C ± 0.5°C, using Apparatus 2 with a paddle rotation speed of 50 rpm. Sampling was carried out at 5, 10, 15, 20, 30, 45 min time points. Analytical signal detection was conducted at (λ) = 240 nm using a Shimadzu UV-1800 UV/Visible Scanning Spectrophotometer (Shimadzu Corporation, Kyoto, Japan).

Data Analysis

Data analysis was performed using MATLAB 8.2 (MathWorks, USA). Principal component analysis and data preprocessing was performed using PLS Toolbox 7.8.2. (Eigenvector Research, USA). Dissolution prediction modeling was carried out using Neural Network Toolbox 8.3 within the same MATLAB environment. Sum of ranking differences (SRD) method was applied using a visual basic applications (VBA) macro in Excel 2023 (Microsoft, USA) made available by the developers (http://aki.ttk.mta.hu/srd/).

Data Collected Throughout the Manufacturing Process

The dataset includes critical parameters from the granulation and tableting processes, as well as spectral data. The granulation process comprised of warm-up, granulation, and cool-down phases. During each phase, the following parameters were recorded at one-minute intervals: inlet air temperature, material temperature, outlet air temperature, inlet air volume, and inlet air humidity (absolute and relative). The average values and standard deviations (SD) were calculated for every parameter for each batch. For the tableting process, nominal lubrication time and compression force data were available. Additionally, transmission and reflection near-infrared (NIR) spectral data were collected from both sides of each tablet.

For each experimental design setting, two lubrication times and three compression forces were applied during tableting, resulting in a total of 90 settings. NIR data and dissolution profiles were obtained for three tablets at each setting, and the data from these tablets were averaged.

Multi-Way Analysis of Variance

The multi-way analysis of variance (ANOVA) was utilized to investigate the influence of experimental factors on tablet dissolution profiles. The analyzed factors included macrogol concentration, granulation time, granulation temperature, lubrication time, compression force, and combinations of these factors. The analysis was performed separately for each time point along the dissolution curve to identify the experimental factors with the greatest influence on dissolution.

Dissolution Prediction Models

Preprocessing and principal component analysis of the input data was performed before training the models. The granulation time-series dataset was subjected to mean centering, which involves adjusting the data so that each variable has an average value of zero. This process ensures that the dataset is normalized, allowing the model to focus on variations in the data rather than absolute values. This was followed by principal component analysis (PCA), where the time-series data for each process parameter were individually transformed, and the first two principal components were retained per parameter to reduce dimensionality while capturing the main variation patterns. PCA applies a coordinate transformation to the original dataset, resulting in new variables that are orthogonal to each other so that the first few components account for highest possible variance in the dataset. This resulted in a total of 12 principal component scores (2 per parameter × 6 parameters), which were then used as inputs to the ANN models.

For the NIR reflection and transmission spectra, various preprocessing methods were tested, including mean centering, standard normal variate (SNV), and Savitzky-Golay first and second derivatives. A partial least squares (PLS) model was used to predict the macrogol content of the granules from their reflectance NIR spectra. This method identifies the underlying relationships between the spectral data and macrogol concentration by projecting the data into a lower-dimensional space. The NIR spectra were preprocessed using baseline correction with the automatic Whittaker filter method (λ = 10,000, p = 0.001) followed by mean centering. The final PLS model with one latent variable, accounting for 91.48% of the variance yielded an RMSECV of 0.5570% w/w and an RMSEP of 0.7410% w/w .

Feedforward-fully connected ANN models were created to predict the dissolution profiles of the tablets. The Bayesian regularization learning algorithm was used as training algorithm which is known to reduce overfitting and improve generalization as opposed to e.g., the Levenberg-Marquard algorithm. This approach minimizes the linear combination of squared errors and weights, effectively limiting the complexity of the models and improving performance on previously unseen data. Hyperbolic tangent sigmoid activation function was used in the neurons of the hidden layer, while a linear transfer function was applied in the output layer. The network architecture included one input layer, with the number of neurons adjusted based on the number of input variables. The optimal number of neurons in the hidden layer was optimized by training the model with a neuron number between 1 and 10 – in each case training was repeated 30 times. The output layer contained six neurons, corresponding to the six points of the dissolution curve. Several ANN models were built with different combinations of input data, including the transmission and reflection NIR spectra, nominal experimental settings, granulation time-series data, and the average values and standard deviations of the granulation data (see section 3.2.1. for details). Nominal experimental settings refer to the theoretical values established during the design of the experimental setup. Data were collected from 90 experimental settings, each represented by the average values of three tablets. Ten settings were randomly selected as independent validation data, while the remaining 80 were used for training purposes. This training dataset was then randomly divided into training, validation, and internal test samples in a 70–15-15% ratio. The training set was randomly divided into subsets, and the initialization of weights and biases initialization algorithm (Nguyen-Widrow layer initialization) also introduced elements of randomness. To account for this variability, a bootstrap resampling technique was employed, consisting of 200 resampling. By this, the reported model results correspond to the mean of the 200 submodels and the 95% confidence interval could be also determined as the 2.5% and 97.5% percentiles of the 200 independent ANN model runs.

The performance of the ANN models was evaluated using the f1 difference factor, the f2 similarity factor, coefficient of determination (R2) and the root mean square error (RMSE). These values were calculated for both the training and independent validation datasets. The f1​ difference factor determines the average percentage difference between two dissolution profiles (Eq. 1):

$$\begin_=50_\left\^\left|_-_\right|\right]/\left[\sum\limits_^_\right]\right\}\times 100\end$$

(1)

where n represents the number of points in the dissolution curve, Rt and Tt​ are the measured and predicted dissolution values at time t.

The f2 similarity factor was calculated using the following formula (Eq. 2):

$$\begin_=50_\left\\sum\limits_^__-_\right)}^\right]}^\times 100\right\}\end$$

(2)

where wt is an optional weighting factor, was not implemented in this study. Only one measurement after 85% dissolution was included in the calculations to ensure that the plateau at the end of the curve does not undeservingly improve the results.

RMSE was calculated according to the equation below (Eq. 3):

$$\beginRMSE=\sqrt\sum\limits_^_-_)}^} \end$$

(3)

For the f1, f2 and RMSE values, calculations were performed for individual dissolution curves, and the results were averaged. In contrast, the R2 values were calculated collectively for all measured and predicted data points (separately for training and validation datasets).

Sum of Ranking Differences

The Sum of Ranking Differences (SRD) method is used in this paper to compare dissolution models for the first time. Proposed by Héberger (46) and later validated and implemented by Héberger and Kollár-Hunek (47), the SRD method provides a non-parametric, robust and generalizable approach for comparison, most often used for comparing various models or methodologies (48,49,50). It evaluates the closeness of actual rankings to an ideal or references ranking by calculating the sum of absolute rank differences (SRDs) for each model/method across all objects compared to the reference ranking. The main steps of the SRD method are summarized in Fig. 2. At each dissolution time point, the models were ranked across the n validation samples. The smallest value is assigned a rank of 1, the next smallest rank 2, and so on, with the largest value receiving rank n. For every sample i (where i = 1,…,n, and n is the number of dissolution profiles), a reference rank ri,ref was assigned based on the measured dissolution values. The corresponding predicted value from ANN model j was ranked in the same way, giving ri,j. The sum of ranking differences for model j was then calculated as

$$\begin_=\sum\limits_^\left|_-_\right|\end$$

(4)

with lower SRD values indicating closer agreement between predicted and experimental rankings, i.e., a better model/method.

Fig. 2

Main steps of the SRD method

These SRDs are scaled between 0 to 100%, 0% indicating a perfect match, while 100% means a completely reverse ranking. The SRD calculation method is also complemented with the generation of a high number of random rankings to obtain a Gaussian distribution, which can be used to assess if the rankings differ from random rankings.

In this study, we used the SRD method to compare the 10 ANN dissolution surrogate models. The calculations have been performed separately for each time point of the dissolution curve, i.e. a standalone SRD ranking order was obtained for each time point.

At each time point, the measured dissolution values of the tablet served as the reference ranking, and predicted dissolution values of the validation tablets were evaluated against this reference. Consequently, the SRD metric in this case indicates if the order of the predicted dissolution curves compared to each other how well follows the ordering of the real, measured dissolution curves. Essentially, this indicates the discriminatory power of the models: a surrogate model demonstrates high discriminatory power if it accurately replicates the rank order of dissolution profiles across varying tablet settings, reflecting its ability to capture the critical factors influencing dissolution.

The dataset was structured such that rows represented the individual samples, while columns represented the predicted values by the different models, and the measured dissolution values as the reference data. For the validation set, there were 10 × 3 tablets, but the dissolution curves of the three tablets were averaged, resulting in 10 average dissolution curves for SRD analysis.