C1 – AI/ML for Clinical Trials and Humanized AI for Future Healthcare

Chair: Hwanhee Hong (Duke) 

Instructor:  Mark Chang, PhD (Boston University)

Course Description: 
Machine-Learning (ML) for Clinical Trials and Humanized AI (HAI) play increasingly critical roles in our work and daily lives. This short course will consist of two parts that are based on the instructor’s two books: Artificial Intelligence for Drug Development, Precision Medicine, and Healthcare (2020), and Humanized AI – Foundation, Architecture, and Prototyping (2023). In the first part, we review current research in Medical AI, the landscape of AI/ML in drug development and healthcare, the challenges we are facing, e.g., regulatory perspectives in AI applications, and we discuss the similarities and differences between classical statistics and AI/ML. Through several motivated examples, we will illustrate the benefits of the paradigm shift from classical statistics to AI/ML in drug development. With implementation in R, we will demonstrate an AI application in rare disease clinical trials. In the second part, we discuss artificial general intelligence (AGI) or humanized AI, its impact on daily life and health, how it will reshape healthcare systems, and how we can guide HAI development in a positive direction. We will survey existing big-data-based HAI/AGI approaches and discuss how new small-data-based approaches can work better and make a more profound impact, while discussing attention, learning, and responses mechanisms and their implementations with HAI demonstrations. This short course is designed for both audiences who just want to learn the overall perspectives of AI and those who are more interested in hands-on work.

We will investigate the similarities and differences between AI and ML and between classical statistics (CS) and ML, as well as the benefits of shifting from CS to AI/ML. We provide an overview of five types of machine-learning (ML) mechanisms: (1) supervised learning, (2) unsupervised learning, (3) reinforcement learning, (4) collective intelligence learning, and (5) evolutionary learning. Natural-Language (NL) text is one-dimensional, while images are two-dimensional, and motion pictures and molecular structures are 3 dimensional. However, such high dimensional structures are translated into one-dimensional information, and thus the same deep learning architectures and methods can be applied as those in the NL process (NLP). We will review the basic deep learning architectures, including (1) Feedforward Neural Networks (FNNs) for general classification and regression, (2) Convolution Neural Networks (CNNs) for image recognition, (3) Recurrent Neural Networks (RNNs) for speech recognition and natural language processing, and (4) Deep Belief Networks (DBNs) for disease diagnosis and prognosis, and (5) Generative Adversarial Networks (GANs) for classification problems, (6) Autoassociative Networks (Autoencoders) for dimension reduction, (7) Deep Belief Networks (DBNs) for cardiovascular risk prediction, and (8) Variational Autoencoders (VAEs) for probabilistically generating new data. We’ll discuss developmental relationships among these deep learning architectures as well as more recently developed deep learning models, including Generative Pre-trained attention transformer architectures used in NL models such as chatGPT. AI/ML methods have been widely used in target identification, molecular design, disease diagnosis and prognosis, and healthcare in general, but make limited use of the big-data-based ML approaches in clinical development or clinical trials due to small data and other issues. We will discuss the applications using R as a software example. There will be a discussion of the challenges, including regulatory aspects, and how to bring ML into clinical trials and clinical development programs with examples.

Later in the short course, we will discuss a Humanized AI (HAI) that is far beyond the capabilities of NL models such as ChatGPT. We demonstrate how such a “futuristic” model is coming in the near future. HAI involves making agents who look, think, and behave like human beings, and in a broad sense are machine-race humans who can serve as our (especially seniors’) physical and soul companions. We will discuss the key concepts in HAI such as the connotation of understanding, self-awareness, consciousness, morality, creativity, imitation, imagination, and cognitive learning. We’ll discuss building architectures that include attention, learning, and response mechanisms. We will demonstrate how the mechanisms can be implemented based on the similarity principle, hierarchical tokenization and recursive patternization, and how to implement HAI, e.g., using Swift in Xcode. Of course we will also discuss the controversies of building such super AIs