Artificial intelligence (AI) tools in medicine are rapidly advancing in technical ability and are increasingly being relied on by physicians to guide clinical decisions. In this chapter, we provide an overview of AI tools in oncology, with several examples of state-of-the-art diagnostic and therapeutic tools. We follow this overview with a discussion of legal liability for both AI manufacturers and physicians when AI tools are used in the course of patient care. Next, we discuss ethical implications of AI algorithm design and use in clinical care. We end with a discussion of how AI tools promise to shift financial risk for various health care stakeholders.
ARTIFICIAL INTELLIGENCE IN ONCOLOGY
Oncology is a branch of medicine that focuses on the prevention, diagnosis, and treatment of tumors. Oncologists traditionally work with radiologists and pathologists and manually read images and clinical notes, a process that is susceptible to individual subjectivity resulting in delay and diagnostic irreproducibility. In contrast, AI takes advantage of big data, automates classification and prediction, and thus provides early detections that may be invisible to the human eye and treatment recommendations that incorporate the vast volume of medical literature and comply with up-to-date guidelines.
AI is the branch of computer science concerned with the automation of intelligent behavior. It refers to intelligent machines (computer algorithms and robots) that think or act rationally or human-like.1 In fact, AI has evolved from being mathematical logic based during the 1960s to 1990s to statistical learning based (machine learning [ML]) since the 1990s and has advanced since 2012 with new methods such as deep learning (DL).2,3
The mechanism of AI is a function, mapping from an object to a class to which the object belongs. This process can be visualized as a machine whose input can be an image of an object (e.g., magnetic resonance imaging [MRI] scan of brain tumor) and whose output is the classification of that object (e.g., a predicted label of the tumor as cancerous or not). The machine works through three main steps: First, extract and refine features of the input. Second, select a classifier (e.g., decision tree, naive Bayes, maximum entropy, hidden Markov model) that analyzes those features and produces the class/label to which the input belongs. Third, evaluate the performance of the classifier.4 Based on the one-layer mapping described earlier, DL adds additional layers of mapping—artificial neural networks (ANNs)—to achieve better performance.
Input consists of two types: (1) structured data in electronic health records (EHRs) that are entered into drop-down boxes and point-and-click fields, and (2) unstructured data in narrative forms in clinical notes such as written case reports taken from consultations.5 Depending on the type of input, AI has different areas including, but not limited to, (1) image recognition that processes and classifies pictures, (2) natural language processing (NLP) that works on texts, and ...