Data acquisition in pharmaceutical production processes and its use for quality predictions

Learn about the value, capture, quality predictivity and use of data to optimise productivity in the pharmaceutical industry. Use and collection of data in the pharmaceutical industry, using sensors as a valuable source of information to monitor and control critical parameters of the production process.

The trend towards extracting value from data, its capture and reliability in industrial optimisation has taken off, providing precision, prediction and quality in final products, and the pharmaceutical industry is no exception. Pharmaceutical production stands out for innovation and development aimed at saving lives, recognised for its high quality standards, although rigorous when implementing improvements in production, due to the demanding regulations it experiences.

Regulatory bodies such as FDA and EMA are inclined towards the incorporation of data-driven manufacturing, stimulating digitisation, recognising that the acquisition of data in pharmaceutical production processes and its use in quality predictions are favourable. The use of Data Science allows the digitisation of production lines, reducing downtime, improving the management of failures, reducing waste, among other benefits.

Sensors for data acquisition in pharmaceutical production processes to produce medicines

Sensors used in the actual production of pharmaceuticals serve to monitor and control the critical parameters of the production process, they are a source of data describing each phase, from incoming raw materials, process quantifications, intermediate product characteristics to the quality of the final product. Harnessing the value of data in the pharmaceutical sector and using it predictively for final quality requires working with the range of data obtained, from laboratory analysis of incoming raw materials to complex time series results that can be used for every second of manufacturing.

Valuable data provides serious insight into the process trajectory compiled over the years. A collection of big data applied in the large-scale production (1005 batches) of a blood cholesterol-lowering medicine, whose dosage form is film-coated tablets, under an immediate-release drug profile, composed of excipients (lactose, silicified microcrystalline cellulose and starch), which underwent pre-use quality control analysis and an active pharmaceutical ingredient (API).

The information inputs generated every second of the process are stored in a SQL database on the tablet press, then uploaded to a server for expert viewing or extraction. Semi-structured data that requires cleaning and organisation before use.

The capture of the data set corresponding to the product family composed of several sub-families, are differentiated in strength and manufacturing batch size. Distributed in four different concentrations and nine different batch sizes. Data collected for more than one year, allowing to consider seasonal variations, changes in raw material input, impacts of shift work of operators, holiday breaks; ensuring the robustness of the Big Data (macro data) collected in the manufacture of the selected product.

The main source of data comes from results of laboratory analysis of incoming raw materials, tablet core and final product HPLC (high performance liquid chromatography), gas chromatography, moisture analyser and particle size analyser up to automatic tablet core analyser.

Another source of data is time series applied in tablet compression, the output of time series such as: tablet press speed, compaction force, filling depth, etc.

Batch genealogy data 

Extracting value in the capture and reliability of data in quality pharmaceutical production requires an adequate organisation, prior to accessing and safely exporting the inputs obtained through the sensorisation of the production line, the so-called batch genealogy is carried out. A rigorous recording of laboratory and process data stored in databases with their descriptors.

Subsequent to this initial sweep, collection is triggered, grouping specific information by manufactured batches under material codes and product sub-families. The export filter settings, including time intervals, product code and laboratory analysis range.

The challenge in data collection is the extraction of time series from the manufacturing process, compared to laboratory data due to their quantity. These are periods that can contain between 2 and 20 hours, discriminated by product code, batch size (tablets produced), collected every 10 seconds, allowing to reduce the complexity and maintain the reliability of the information.

Data pre-processing phases

The use of data in the production of medicinal products and the predictive nature of their final quality, contemplates pre-processing phases of information, both from the laboratory as of time series.

The stage of the laboratory examines the results of the analysis of the incoming raw materials, the quality of the intermediate product (independent variables) and the quality of the final product (dependent variable), finding outliers. It assesses whether the data have the expected spread based on product and process knowledge. Under the parameters described below:

1. Batches without final product quality results are excluded.
2. Four different product sub-families with different central weights of target tablets lead to the creation of a standardised parameter (RSD) with a relative standard deviation of the weight.
3. Each product code (sub-family) has a different thickness, target, diameter and hardness, creating a new standardised parameter for tablet dimensions and shape: tensile strength. Since this is comparable between the subfamilies of the product, because they have the same shape, the equation is not modified. A new parameter was created by applying the following equation:

Where F represents the hardness of the tablet in Newton (N), t is its thickness in millimetres and d represents the diameter in millimetres. The calculation considers the average hardness value and the maximum thickness and diameter for both tablet cores and film-coated tablets.

5. The product quality parameters included in the dataset are final product impurities, residual solvents and drug release results.

In the quality of data on time series, unusual events and requirements for special pre-processing or elimination of certain batches are observed. The parameters included: the speed of the tablet press (indicator of when production starts), and the number of rejected tablets counted by the press, (provides information on the speed of the tablet press, that if it reaches 0 the process is finished, there was another start-up or there were problems) were the most descriptive of the batch dynamics. The following pre-processing and cleaning steps were taken:

1. Weekend stoppages, although a value of 0 was present, due to the pause of the press, this feature is deliberately considered as part of the data set, because leaving the mixed powder material at rest could affect the quality of the tablets.
2. Standardisation of the time format, by including unstructured data in the SQL database of the tablet press, and transferring it to an available database (iHistorian), showing a mixed time structure of 12 and 24 hours.
3. A drop in the parameter of rejected tablets in some batches leads to investigate particularities, discovering problems extending over two or more days, due to incorrect date structure in some batches, and correcting them.
4. Detection of the so-called “Gap” in some series, indicative of several minutes without available data. Motivated by shutdown in the tablet press (shift changes, press calibration, calibration of the automatic weight and hardness control system, etc.).
5. After the above steps, the time series is corrected and the data is cleaned, performing another cycle of pre-processing checks in perfect operation.

Creation of new data parameters 

The data captured in the pharmaceutical process is used to predict the quality of the finished product, so the entry of laboratory data for each batch describes the time series and its specificities, giving rise to the combination of new parameters that deepen further analysis in production and quality prediction. Among the data in reduced format as new attribute vectors, adapted separately by manufacturing parameters based on expert knowledge, the following stand out:

Average tablet press speed, The "longest interruptions" (weekends and public holidays), which mask the real characteristics of the process, are known.

• Number of changes in tablet press speed, The "normalisation" of the batch size allows the batch size to be standardised, longer batches naturally have more interruptions than smaller ones.
• Tablet press speed of 0, creation of an additional categorical attribute indicating execution during weekends, adjusting the time of interruptions.
• Total number of tablets rejected, The cumulative quantifier parameter for each time point in the process of rejected tablets. It is normalised to the batch size.
• Total number of tablets rejected at start-up, Indicator of time and effort spent in preparing the tablet press parameters for a particular mix. Provides information on the most challenging material properties or inexperienced operators.
• SREL average start-up, The average relative standard deviation of the main compression force (SREL), during compression start-up, excludes values where the tablet press does not operate.
• Average SREL, The average SREL during the compression run.
• SREL max, maximum SREL value during a compression run. It is essential to eliminate excessive values because they are not realistic data.