Artificial intelligence in pharmaceutical domain (with emphasis on the data quality). ALCOA prediction from pharmaceutical industry line

Abstract

We can all imagine the amount of data generated during the procedure of the production of a medicine in a pharmaceutical industry. Data are taken from the import of the raw material to the factory, its analysis until it is used in the production of the drug, from the production line, from the warehouses but also from the distribution lines up to the final consumer, the patient. All this data must ensure traceability, if possible, from the raw material production plant to the final consumer. The most recent example is the example of the pandemic of COVID 19 vaccines. Therefore, before a drug can be used, the pharmaceutical company must prove that the drug is effective and safe. That is why the pharmaceutical companies are conducting many tests, and numerous studies in quality control. Over the years, the pharmaceutical companies have adopted the concept of ALCOA as a framework for ensuring the observance, preservation, security and accuracy of data. The term ALCOA is an acronym that means Attributable, Legible, Contemporaneous, Original and Accurate. From the meaning of these words, we can easily understand why this acronym is so important and why it was adopted by the pharmaceutical industry. The term ALCOA is about the quality and integrity of the data, which has a direct impact on the quality of the drug. This master thesis is an attempt to classify the values obtained from different sensors (from two production lines of a well-known Italian pharmaceutical company) as data that are Attributable and Contemporaneous. This attempt was done by using three deep learning models. We also tried to find out if there is a possibility to predict the next ALCOAs from the previous ones. These words came from the acronym ALCOA that mentioned above. The three deep learning models used are the LSTM Model, Bi-LSTM Model and GRU Model. Unfortunately, the above deep learning models failed to predict the next ALCOA the Attributable and the Contemporaneous from the previous ones. The three models used showed better performance in the Attributable than in the Contemporaneous yet again this performance does not allow us to use them as models for predicting this ALCOA in a pharmaceutical industry. However, this should not disappoint us as it is the first attempt to use such models in the prediction of ALCOA. After all, this dissertation focused only on the Attributable and the Contemporaneous. There are three other letters in the acronym as well and the dataset was only from two production lines. There are so many deep learning models, machine learning algorithms, so many more letters remaining letters in the acronym, other and more improved datasets that can only give promise for the future.