Drug Evaluation Committee Machine learning has begun.
May 2020
This report uses actual data to explain the processes of planning, data processing, and machine learning implementation, with a focus on scenarios where data scientists—in the broad sense, including those working in the pharmaceutical industry—apply machine learning and AI to develop services using readily available data, such as clinical data.It also introduces methods for increasing the volume of data to compensate for the lack of training data—a major obstacle in AI development—as well as approaches to deep learning that utilize pre-trained models. Due to the difficulty of securing sufficient data, this report does not cover deep learning from scratch.
“Chapter 1: Projects Involving Machine Learning” provides an overview of the differences between using machine learning and conventional system development.“Chapter 2: Acquiring and Processing (Annotating) Development Data” and “Chapter 3: Data Partitioning and Training Methods” explain how to acquire and augment data, how to process it, and how to partition it for machine learning, including the methods used in Chapters 4 and 5."Chapter 4: Implementation Examples of Applying Machine Learning to Clinical Trial Data and RWD" provides a detailed account of the process—from actually performing machine learning using data such as blood glucose levels of diabetes patients using insulin to deploying the results in a service.Furthermore, “Chapter 5: Explanation and Practical Examples of Using Pre-trained Models” introduces the concepts and practical applications of transfer learning and fine-tuning—which involve modifying and reusing pre-trained models—using examples from image and natural language processing. The program code for the machine learning examples presented in this book is included.While these are not the most cutting-edge methods or case studies, the abundance of open-source libraries and other resources, along with the wealth of reference information available online, makes them accessible. Interested readers might want to start with something simple and familiar—such as performing logistic regression analysis using a single-layer neural network.We hope this book will serve as a helpful resource for users in the pharmaceutical industry—particularly those in clinical development and post-marketing settings—who wish to utilize machine learning techniques.
Japan Pharmaceutical Manufacturers Association (JPMA) Drug Evaluation Committee, Data Science Subcommittee
Task Force 1 Machine Learning Team