AI tools could transform how brain tumor patients access and understand critical care information, but without careful oversight, the same technology may introduce new risks and uncertainties.
Study: Large language models in patient education for brain tumors: opportunities, risks, and ethical considerations. Image credit: Nanci Santos Iglesias/Shutterstock.com
Patients with brain tumors suddenly need to understand a lot of information related to their condition and to medical care, coupled with emotional struggles and cognitive overload. A review in Frontiers in Oncology concludes that large language models (LLMs), when properly supervised, can be useful tools for improving patient understanding and involvement in their care.
Brain tumors overwhelm patients with a sudden cognitive and emotional burden
Brain tumors are life-changing for both patients and their families, often appearing suddenly with alarming symptoms such as seizures or cognitive impairment. As the disease progresses, it can lead to personality changes, memory loss, or paralysis, compounding both emotional and functional distress. This burden is further intensified by poor outcomes; for example, glioblastoma carries a five-year survival rate of less than 10 %.
Despite the existing strain, these patients and their families need to be educated about the disease, the type of interdisciplinary care involved, the risks of each therapeutic approach, the prognosis, and the support available. In many cases, these are individuals without significant health literacy.
The current patient-oriented literature on brain tumors typically requires at least a high-school education, and often more, limiting its accessibility to those who most need it. Physicians often explain things well, but there is too much to take in at one time, and consultation time is limited. Anxiety and cognitive overload make it difficult for patients and caregivers to understand and remember this important information, and to get new information as the condition changes. Unable to get the answers they seek, they go online or join support groups.
Aware of the challenges of providing patient education in a manner that satisfies patients and their families, the authors of this paper examined LLMs for their ability to fill this gap. As a narrative review, it selects themes considered important by the authors, perhaps introducing selection bias. Their choices were directed by experts, however, based on a curated body of literature.
LLMs are artificial intelligence (AI) systems trained on large amounts of data to provide human-like answers, simplify when asked to do so, and clarify in specific situations. They can handle multiple chats simultaneously, unlike healthcare providers, who can see only one patient at a time and a limited number per day.
AI tools may support understanding, but lack true clinical insight
LLMs are trained to respond politely and with reassurance, conveying a sense of empathy. This could provide emotional support to the distressed patient, though evidence of sustained real-world impact remains limited. LLMs can be integrated with other platforms to explain complex procedures, test results, and the effects of various treatments at an individual level. This allows patients to potentially feel heard and supported. LLMs can also provide ongoing patient guidance outside the treatment setting, thereby strengthening medical advice.
Overall, they may provide patients with much-needed, clear, and relevant answers to general questions about their diagnosis and treatment, and potentially at the right time. They may still produce outputs that are too technical or advanced in reading level without careful prompt design.
LLMs can be very useful for explaining preoperative cognitive tests to such patients. These tests are key to planning surgery, but require a lot of time to explain. They could, though not yet consistently, convert structured radiological reports into understandable explanations in real-world neuro-oncology settings. LLMs currently do poorly at interpreting sophisticated neuroimaging results, such as magnetic resonance imaging (MRI), and apparent successes often involve explaining radiologist-authored reports rather than directly analyzing raw imaging data. They may attempt to simplify such reports, which can lead to misinterpretation and, in some cases, raise concerns about data privacy.
In evaluating LLM performance, the conventional metrics include accuracy, completeness, conciseness, and the safety of information provided. However, the authors note that other qualities should also be assessed, such as readability, cultural appropriateness, anxiety, and empathy, in addition to usability.
Privacy, accountability, and bias challenge safe clinical deployment
Despite the promise of LLMs in patient education, there are serious potential disadvantages to their use. They produce responses to medical queries based on statistical analysis and computational manipulation of the data used to train them. This can lead to inaccurate or nonexistent information (“AI hallucinations”) being given, for instance, about treatment or outcomes. To minimize this, recent research has focused on retrieval-augmented generation (RAG), in which LLMs are constrained to preselected knowledge sources.
In addition, LLMs provide fluent and seemingly authoritative answers that could induce overtrust among patients, potentially obstructing shared decision-making with clinicians. They may also lead to emotional bonding followed by later disappointment when expectations go unmet. These aspects remain poorly researched, though vital to their use as patient-facing tools.
Despite the apparent empathy, AI systems lack true insight and accountability, raising ethical concerns. This may translate into impersonal care recommendations. Patient privacy is another critical area of concern.
Default LLM outputs are generally at an undergraduate reading level or higher, highlighting the need to frame prompts appropriately. This requires clinician training in the use of LLMs.
The way LLMs arrive at their conclusions is difficult to interpret, especially with more advanced multimodal systems that use both visual and textual data simultaneously. The authors suggest that these should be applied carefully in clinical situations. Given their basic probabilistic design, they tend to prioritize comprehensive coverage over strict clinical reasoning. This makes them prone to unwarranted extrapolations and inferences, and variable outputs. LLM output must be verified by neuro-oncologists at the level of decision-making information, such as tumor characteristics and other possible diagnoses, as wrong answers here could increase the patients’ distress.
This emphasizes the need for diligent oversight, transparent outputs, technical guardrails such as RAG, and clinician verification, thereby balancing the advantages of this novel platform with appropriate safety measures. An example of emerging regulation is the Prof. Valmed system. This is an early clinical decision-support tool that has gained EU (European Union) Medical Device CE approval. This heralds the formal regulation of these tools in healthcare. The EU is moving toward mandating LLM use within Human-in-the-Loop architecture, a framework that ensures LLMs act as assistants and not agents in their own right.
Other pressing needs include using better models trained on better datasets. A safe framework for integrating LLMs into clinical practice would involve multiple areas:
Define the intended use
Set clear boundaries
Use structured prompts and mandatory uncertainty disclosure statements
Ensure readability
Make clinician validation mandatory
Secure patient portals to ensure data privacy
Establish safety metrics, including hallucination thresholds and accuracy targets
Train clinicians and patients in safe AI use
The legal responsibility for LLMs in brain tumor patient education might encompass three domains, according to these authors: manufacturer accountability for system performance, institutional responsibility for regulating system implementation, and clinician accountability to validate the final decision.
Safe integration requires oversight, regulation, and improved models
LLMs have the potential for much use in educating brain tumor patients, but future research is essential to validate their outputs across tumor subtypes, especially when the tumor has a poor prognosis or is relatively uncommon. Evidence to date varies across tumor types, with more data available for some (e.g., pituitary adenoma and meningioma) than others, and several subtypes remain underexplored.
Interactions between patients and LLMs also need to be studied, including patient understanding, anxiety, decision-making, and overdependence. Robust real-world validation of patient outcomes remains limited, and enhanced health literacy, the refinement of multimodal LLMs, and accountability remain important goals for the future, helping limit LLMs as assistants rather than autonomous tools in current practice.