Artificial intelligence in pharmaceutical domain (with emphasis on the data quality)

Abstract

Over the past decades, computer science has been widely developed. From speech recognition to reconstructing brain circuits to even natural language understanding, one can easily claim that Machine Learning and Deep Learning have become part of everyday life. They have been used in most of the industries and the pharmaceutical one is no exception. During the recent years, data integrity has been an important part of the pharmaceutical industry. It is mandatory to ensure the quality and the safety a drug has from the start of its production till it reaches the customer. In order to achieve this, the American Food and Drug Administration (FDA) introduced the ALCOA principles as guidelines for every pharmaceutical company. The ALCOA acronym stands for Attributable, Legible, Contemporaneous, Original, and Accurate which are values derived from the pharmaceutical industry data collected during the manufacturing process and which ensure the integrity of the data. SPuMoNI is a European funded research project that explores the authenticity of such data using innovative scientific approaches. This thesis presents, an experimental attempt to predict the ALCOA values while using the raw sensor data as an input. To achieve the best regression result, three Deep Learning Recurrent Neural Networks have been used. Long Short-Term Memory, Bidirectional Long-Short Term Memory and Gated Recurrent Unit. More specifically, there are two ALCOA values that are being examined in this thesis, Legible and Accurate. The language used for programming the algorithm was Python through Google Collaboratory and the library imported for the deep learning methods was Google’s TensorFlow. Through the research conducted, it was shown that the prediction’s accuracy of the ALCOA principles Legible and Accurate was below 80% which drives us to research a) which additional parameters need to be taken into account and b) which algorithms such as a Transformer Neural Network could lead to accuracy improvements.