Promise and Perils of Large Language Models for Cancer Survivorship and Supportive Care

Abstract

A call to action to bring stakeholders together to plan for the future of LLM-enhanced cancer survivorship.

Introduction

Cancer symptom toxicity affects patients and caregivers of all cancer types and stages, with effects on physiologic, psychologic, financial, and social well-being.¹ Barriers to symptom assessment and management of patients with cancer include lack of survivorship care access and resources, incomplete data for risk-stratification, inadequate means for timely communication about symptoms, fragmented health care across multiple specialists, misconceptions and insufficient patient and clinician education, and changing needs over the cancer trajectory.^1-7 Large language models (LLMs) are uniquely poised to address these barriers because of their ability to process, transfer, and transform high-dimensional knowledge, providing more individualized management and enabling multilevel communication that can adapt to users' needs. Here, we discuss the current status and future applications of LLMs for cancer symptom management and call for cross-disciplinary collaboration that centers the needs of patients and caregivers. Table 1 provides an overview of key opportunities and existing challenges.

TABLE 1.

Key Opportunities and Existing Challenges for LLMs in Cancer Survivorship and Supportive Care

Key Opportunity	Example	Challenges
EHR-based symptom/adverse event collection	An LLM fine-tuned to identify adverse event mentions in clinical notes is used to build a large supportive care and survivorship registry	Underreporting of symptoms/adverse events in EHR Fragmented health care and data siloes Ambiguities in definition and description of symptoms and adverse events
Outcome prediction	An LLM fine-tuned to predict the occurrence of severe platinum-induced peripheral neuropathy provides risk estimates for individualized shared decision making about treatment options	Symptom/adverse event outcomes are needed to fine-tune LLMs but are underdocumented Institutional buy-in for operationalization and monitoring needed for successful implementation Lack of interpretability into how LLMs arrive at their prediction Algorithmic bias arising from historic differences in outcomes and access to care across population Predictions could be used by other parties not in the service of patients (eg, denying insurance coverage)
Access to survivorship care	LLM-enabled EHR automatically identifies billing codes and drafts notes, reducing clinicians' administrative burden and creating more time for them to see patients	LLMs hallucinate and have risk of harmful language/bias learned from on pretraining data Lack of automated metrics to evaluate quality of LLM performance over time Risk of over-reliance on LLM output, automation bias, and clinician deskilling
Remote symptom monitoring and data collection	An LLM chatbot proactively asks patients about symptoms on the basis of the treatments they have received. The chatbot adapts to patient responses and asks clarifying questions for high-quality reporting. Results are fed back to the EHR for clinician review	Existing out-of-the-box LLMs often provide incorrect recommendations and are not ready for clinical use Inequitable access to technologies (eg, computers, phones, broadband Internet, cellular service) Careful protections needed to protect patient privacy for remote data collection and transfer EHR vendor and institutional buy-in needed for integrating data from external apps Not all patients and caregivers may want to interact with LLMs for their cancer care Digital remote care could place more burden on patients and caregivers LLMs hallucinate and have risk of harmful language/bias learned from on pretraining data LLMs are sensitive to small changes in how they are prompted Lack of task-specific, patient-centered evaluation metrics for longitudinal monitoring
Remote supportive care and survivorship management	An LLM chatbot answers patient's questions about managing symptoms from home and providers appointment reminders. High-risk responses are triaged for urgent review by a clinician
Patient and clinician education	An LLM chatbot provides a user-friendly interface to identify relevant supportive care and survivorship guidelines	Could shift care burden from healthcare institutions and clinicians onto patients and caregivers LLMs hallucinate and have risk of harmful language/bias learned from on pretraining data LLMs are sensitive to small changes in how they are prompted Lack of task-specific, patient-centered evaluation metrics for longitudinal monitoring
Patient-clinician communication	An LLM translates clinical language into simpler language or other forms that are most preferred by each patient
Patient navigation	An LLM summarizes recommendations from different specialists and future appointments

Abbreviations: EHR, electronic health record; LLM, large language model.

What are LLMs?

Natural language processing (NLP) is a field of artificial intelligence (AI) that enables computers to understand human language. NLP methods range from simple rules-based systems to statistical models to the current generation of machine learning models on the basis of deep neural networks.^8,9 Language models refer to deep neural networks trained on vast amounts of text to process human language. Since 2018, a type of neural architecture known as transformers has achieved state-of-the-art NLP performance.^10,11 Using previously unavailable computational power, transformer-based language models have been trained with increasing text data curated from the Internet, leading us to the current LLMs that are orders of magnitude larger than their predecessors, exemplified by ChatGPT.^12,13

These newest LLMs are a type of generative AI, which refers to AI methods that create new content. LLMs have received most attention for their ability to carry out generative tasks, for example question-answering (eg, chatbot-type applications), summarization, and translation. LLMs can pass the US Medical Licensing Examination,^14-17 provide empathetic responses to medical questions posted on Reddit,¹⁸ and translate biomedical and clinical jargon into more accessible summaries.¹⁹ LLMs can also be used for nongenerative tasks such as predicting outcomes²⁰ and extracting information from text for automated real-world data collection.²¹

Potential of LLMs

There is a significant body of NLP literature on developing methods to automatically extract symptoms and adverse events from the electronic health records. Previous work relevant to cancer survivorship has used smaller language models to automate detection of provider-reported adverse events,^22-24 symptoms,²⁵ and patient-reported outcome characterization.²⁶ Developing these models historically required a large quantity of manually labeled data to fine-tune the model for specific adverse events or symptoms—an expensive, expertise-intensive task. Better performing LLMs may require much less labeled fine-tuning data for this type of information extraction. We therefore anticipate LLM-driven advancements in our ability to automate electronic health record–based symptom surveillance. This could benefit patients at the individual level because their symptoms could be tracked more thoroughly, allowing more timely and precise management.^8,9 Automated collection of symptoms and adverse events will enable larger, more comprehensive postmarket surveillance studies, reduce the resources required to track symptom toxicity on trials, and provide real-world evidence to better understand risks, associations between symptoms, and management strategies. However, symptoms are underreported in the electronic health record, and better collection of patient-reported symptoms is needed to realize this promise. As we discuss below, LLMs could also improve collection of these data.

An emerging application of LLMs for cancer symptom management is in developing models to predict outcomes. Recent work has shown that LLMs trained on clinical documents can predict in-hospital mortality, comorbidity index, 30-day readmission, and insurance denial prediction.²⁰ Symptom risk prediction could support treatment decision making and personalized monitoring schedules by providing more individualized estimates of the risks and benefits of treatment. Multimodal models that combine clinical text information with other types of health data, including imaging, structured data, patient-reported outcomes, and patient-generated health data is a promising area of active research.^27-32

LLMs may be most transformative in enabling patient-centered remote supportive care and symptom management and data collection, which are essential to improving cancer outcomes.^33-36 Some studies have demonstrated a quality of life and survival benefit of digital health solutions for symptom monitoring and management, although results vary across studies and specific outcomes.^35-37 Studies incorporating digital health solutions with AI have focused broadly on education, behavior change, and symptom management, particularly pain, anxiety, and depression. Patients report improved empowerment and communication with health care providers and real-time symptom monitoring when using these technologies.

LLMs could facilitate personalized interaction on the basis of up-to-date guidelines and proactive collection of patient-reported outcomes. Compared with chatbots for symptom collection, chatbots that help patients self-manage and triage are higher risk of harm because they directly affect management³⁸ but could potentially reduce health care utilization and help patients avoid urgent care. Key opportunities for improvement include uniform standards for measuring patient-related outcomes, consistent research methods across studies, and integrated patient and clinician involvement.³⁶ Well-designed studies are still needed to show effectiveness of LLMs in symptom management and usability.

The inherent strengths of LLMs in transforming language for summarization and translation create a promising foundation for automating, scaling, and personalizing patient-centered digital health. Methods that summarize and simplify our lengthy, jargon-full guidelines that are difficult even for experts to manage could be a major step forward for patient education and could be tailored to patient preference and background. Automated LLM summarization of clinical visits could potentially improve the longstanding challenge of communication between different clinical specialties, patients, and families—although data silos will need to be overcome for significant advances in this area. LLMs are excellent at machine translation, which has clear implications for improving access. LLMs and similar models are advancing to handle multiple data types,^27-31 providing an avenue to further personalize content to how each patient prefers to consume it (eg, in writing, video, audio, infographic). Similarly, patient- and clinician-facing LLMs could improve education by navigating emerging evidence, guidelines, and management strategies in the context of each patient's history—all of which is increasingly important in our ever-evolving complex cancer care landscape.

Perils of LLMs

However, the promises of LLMs will only be realized if their harms are effectively studied and mitigated. Several technical, clinical, structural, and ethical challenges need to be addressed for effective deployment for cancer symptom management. Most importantly, patients and caregivers should contribute to LLM governance to prioritize ethical, sustainable development and implementation.³⁹ Notably, patients' voices have been largely excluded from the AI conversation. Patients may experience real safety risks, including physical, emotional, and financial harm.⁴⁰ This is especially concerning as it is the patients' data that are needed to develop and refine models, and their health care and privacy are most at risk with these applications. Patient-led advocacy groups such as The Light Collective⁴¹ are exploring ways that patient-centered governance of technology can be operationalized. In collaboration with stakeholders, more regulatory oversight from national and governmental organizations is needed to guide and safeguard implementation.⁴²

Patients should provide foundational input on how and whether they want LLMs integrated into their care and on how they want their health data and public data used to develop models. Transparency is paramount—patients and the public in general should be made aware of how their digital data are used. Especially given the unique ability of LLMs to mimic human interaction, it should also be made clear to users when they are interacting with AI, and informed consent should be obtained as appropriate.³⁹

Patients and their caregivers will have varying levels of interest and aptitude in learning and using this technology. New advances should expand and integrate—not replace—avenues for patients to engage in their care, including the option to opt-out of LLM-enabled digital health. Patients already need to manage and access multiple different applications and patient portal systems during their care. New digital health technologies should simplify systems and not add yet another complication to an already burdensome digital landscape. A few studies are beginning to ask important questions about overall acceptability of such advances including use of AI platforms to monitor distress.⁴³

Patient engagement also requires aligning goals so that their health data are not exploited for the financial gain of developers and institutions while patients continue to face escalating cancer costs. Nefarious use and those not aligned with patient needs and wellbeing need to be anticipated. For example, predictive models could be used by payers and other parties to limit access to care for certain populations if adequate protections are not in place.

Technical issues also need to be addressed for effective clinical translation. One challenge arises from the fact that many LLMs are trained to give a response that a human wants to hear, not necessarily one that is factually correct. While this may be okay for casual interaction, it could be dangerous for health care—especially because LLMs can sound very convincing.^44,45 Emerging evidence shows that while AI chatbots provide cancer treatment information that superficially appears reasonable, results are sub-par when compared held with clinical standards.⁴⁶ How to ensure LLMs are updated as guidelines evolve is an open question and will require close collaboration between professional societies, researchers, and AI developers. LLMs can be sensitive to how they are prompted to give a response, and small changes in the phrasing of a question can yield vastly different responses.⁴⁶ How to obtain consistent and factually correct information from generative LLMs is a research gap. Maintaining a human-in-the-loop will be an important safety check until there is clinical evidence that technologies are safe and effective without additional oversight.

As clinical translation proceeds, all stakeholders will need to work together to design evaluations. Importantly, there is a need for new standards for evaluating generative output. Computed metrics for generative text do not meet the bar for clinical acceptability,^47,48 and patient-centered outcomes and user preferences should remain the standard by which we evaluate. There are no benchmark clinical data sets for cancer survivorship and these are urgently to ensure that best methods are prioritized for translation. In addition, federal and institutional regulatory guidance for longitudinal performance and usability monitoring will need to be carried out on these models given the inevitably of performance drift over time.⁴⁹

While LLMs could democratize cancer care by making self-management and patient education resources widely available, they may also amplify cancer inequities and biases. Like all AI, LLMs learn the biases, racism, and prejudices present in the data that they are trained on.⁵⁰ For example, LLMs could perpetuate ongoing inequities by providing differential recommendations on the basis of previous learned distributions of treatments, monitoring, and management practices across patient populations. In addition, technologies are not equally accessible and often require access and ability to use computers, broadband Internet, or cellular services.^6,51,52 Investments in LLM technologies must be balanced with investments in access for those who need it most.

Our Call for Collaboration and Multidisciplinary Efforts

Cancer survivorship and supportive care is founded in patient-centered communication and education, uniquely poising LLMs to help advance the field. Patients, clinicians, and developers should each play a role throughout the development to deployment pathway. Systems should be designed around the needs of patients with cancer and not around the data that are most easily available. Clinicians and developers will need to work and continuously touch base with patients, including a wide and diverse range of voices, to define and prioritize efforts that align with these needs. Developers can help guide data and model design decisions that affect downstream performance and risks while clinicians and patients should provide insight into data availability, acceptable performance thresholds and tradeoffs, and appropriate evaluations. All stakeholders have critical experience and expertise to contribute to cocreating structures that govern LLMs.

Holding ourselves to high standards for evaluation, safety, and ethics will help avoid early missteps leading to medical errors and loss of patient and clinician trust in LLMs. Now is the time to bring all the stakeholders together as we envision the future of personalized supportive care and symptom management enhanced by evidence-based integration of AI approaches to deliver holistic and equitable care.

AUTHOR CONTRIBUTIONS

Conception and design: All authors

Financial support: Danielle S. Bitterman

Provision of study materials or patients: Andrea Downing

Manuscript writing: All authors

Final approval of manuscript: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

Promise and Perils of Large Language Models for Cancer Survivorship and Supportive Care

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/jco/authors/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).