Artificial intelligence in oncology: Path to implementation

Abstract

In recent years, the field of artificial intelligence (AI) in oncology has grown exponentially. AI solutions have been developed to tackle a variety of cancer-related challenges. Medical institutions, hospital systems, and technology companies are developing AI tools aimed at supporting clinical decision making, increasing access to cancer care, and improving clinical efficiency while delivering safe, high-value oncology care. AI in oncology has demonstrated accurate technical performance in image analysis, predictive analytics, and precision oncology delivery. Yet, adoption of AI tools is not widespread, and the impact of AI on patient outcomes remains uncertain. Major barriers for AI implementation in oncology include biased and heterogeneous data, data management and collection burdens, a lack of standardized research reporting, insufficient clinical validation, workflow and user-design challenges, outdated regulatory and legal frameworks, and dynamic knowledge and data. Concrete actions that major stakeholders can take to overcome barriers to AI implementation in oncology include training and educating the oncology workforce in AI; standardizing data, model validation methods, and legal and safety regulations; funding and conducting future research; and developing, studying, and deploying AI tools through multidisciplinary collaboration.

Keywords: artificial intelligence; deep learning; machine learning; oncology.

FIGURE 1

Data types and sources processed by artificial intelligence. The right column exemplifies commonly used data types that can be processed by artificial intelligence. The left column categorizes the data types into three main areas: patient, medical, and contextual

FIGURE 2

Stage of development and deployment among applications of artificial intelligence in oncology. The location of a topic represents the farthest that topic has come in its development, not necessarily the point in development where all solutions in that topic area have reached. Each topic's shape represents its application within the levels of cancer prevention (circle = primary, triangle = secondary [or tertiary], diamond = tertiary)

FIGURE 3

Statistical biases associated with artificial intelligence (AI) algorithm predictions. AI‐based tools look for patterns of association in the data made available to them; they do not establish causation. The sample of data used to develop an AI algorithm may not represent data from other patients treated in other health systems over time. For example, if most of the data used to develop an AI algorithm came from patients <65 years old treated before 2018, then an AI algorithm may not provide reliable estimates for patients >65 years old treated after 2020

FIGURE 4

Social biases associated with artificial intelligence algorithm predictions. This figure depicts the gap between what we need to show the model (i.e., both factual and counterfactual scenarios) versus what happens when machine learning (ML) is trained on existing data. In this example, an ML model is used to identify oncology patients who require opioids for pain management. When using existing data (i.e., secondary use of data collected as part of routine work), the data reflect not only the association between the patient's condition and opioid prescribing, but rather it reflects this association conditioned on the staff's determination if the patient's complaint of pain is legitimate or not. If the staff's decisions are not uniform (e.g., biased by demographics), then some of the patients who were not prescribed opioids will have the wrong label (“opioids not needed for pain control”), whereas they should have had the label (“opioids needed for pain control”). Therefore, the model will be shown the wrong labels and will learn an erroneous pattern