Welcome back to Navigation Refresh, a recurring, informative feature for novice and seasoned patient navigators alike. In this issue, we cover how artificial intelligence (AI) is impacting oncology and patient navigation.
In December 2024, the President’s Cancer Panel featured a series of roundtables on patient navigation and technology. The panel endorsed patient navigation as a critical evidence-based intervention for cancer patients and their families and sought to understand how technology could increase the reach and impact of patient navigation for populations experiencing the greatest cancer health disparities. Technology can support patient navigators by assisting them with documentation, health communication, and care coordination. These technology supports include electronic health records, patient portals, telemedicine platforms, mHealth applications, chatbots, and AI-assisted care coordination platforms.1
AI is a collection of computer science approaches to extract meaning from large amounts of information. Thus, AI could help patients and patient navigators with basic administrative tasks, like scheduling, as well as more complex tasks such as resource referral. At Montefiore Einstein Comprehensive Cancer Center in Bronx, NY, Moadel and colleagues developed an AI-based virtual patient navigator called MyEleanor to engage with 2400 patients who had not completed colonoscopies in 2022 or 2023. As a result of this intervention, 33% of nonadherent patients (n=789) rescheduled their appointment, nearly doubling the rate of colonoscopy completions for those who were previously no-shows.2
These findings are exciting, since healthcare generally and oncology care specifically face workforce constraints that could be partially alleviated with effective technology. Currently, many health systems do not have the financial resources to hire enough patient navigators to assess and address the needs of all their patients. If chatbots like MyEleanor can streamline rescheduling and even address some barriers to care and coordinate appointments across providers to optimize patient time in the clinic, human patient navigators can be freed up to focus on more complex challenges, such as addressing food insecurity, finding temporary housing, and accessing other community resources. Additionally, AI may support clinicians as scribes, smooth prior authorizations, and help to personalize treatment plans.
If data are not inclusive of the broad diversity of oncology patients seen in our clinics, AI predictions may not be relevant—and may even be erroneous—for subpopulations not included in the original studies.
However, there are concerns about AI in healthcare. AI can only analyze data that it references; therefore, biases that are present in research are likely to be amplified by AI. If data are not inclusive of the broad diversity of oncology patients seen in our clinics, AI predictions may not be relevant—and may even be erroneous—for subpopulations not included in the original studies. For example, when studying the ability of AI models to distinguish between malignant and benign lesions, models that worked well in their original context did not do as well with a larger data set of more heterogeneous skin tones—specifically, the models worked less well for images of darker skin tones.3 Equally concerning, a 2024 study showed that AI large language models tended to stereotype races, ethnicities, and genders.4 In other words, biases that may be subtle in one study can be encoded and amplified by AI in ways that distort the prevalence and presentation of diseases and can lead to biased diagnostic and treatment recommendations.4 A real world example: UnitedHealth was sued for using AI to deny appropriate medical claims to elderly patients due to faulty algorithms.5
AI is only as good as the human algorithms that it encodes. Data sets that are inclusive and generalizable for real-world diversity of patient presentation, lived experiences, and demographics are critical to optimize the usefulness of AI models. Likewise, ongoing clinical feedback from humans to AI to refine algorithms and improve clinical reasoning is critically important. Finally, while the time is ripe for using AI for administrative uses (eg, transcribing meetings, scheduling, revenue forecasting, tracking supply chains), AI should supplement—not replace—human judgment for complex decision-making (eg, providing clinical decision support, optimizing prescription drug recommendations, monitoring patient safety).5 AI can be supervised or unsupervised: Supervised learning ensures that experts oversee the training of the AI model. Unsupervised AI can produce “hallucinations” or inaccurate, misleading, or even nonsensical predictions.5 AI has the potential to increase efficiency and accuracy in healthcare with rigorous clinician feedback and consistent validation checks. Without these important precautions, AI will very likely amplify cancer health disparities that currently exist.
This edition of Navigation Refresh aligns with Standard 1 (Ethics), Standard 3 (Knowledge), and Standard 18 (Practice Evaluation and Quality Improvement) of the Professional Oncology Navigation Task Force (PONT).6
Keep up to date with the latest news from us via social networks:
To sign up for our print publication or e-newsletter, please enter your contact information below.
