Oncologists are being urged to discuss patients’ AI use in consults to help minimise the risks to patient care and privacy rather than discourage the practice altogether.
Writing in a review article for JCO Clinical Cancer Informatics [link here], a group of radiation oncologists pointed out that patients were increasingly relying on AI applications for health information regardless of clinical guidance.
They said this created an “urgent need” for doctors to engage proactively rather than reactively by having conversations on the uses and benefits of AI but also on concerns around the accuracy of information provided and privacy.
The researchers conducted a narrative review on the current landscape and evidence for AI, finding that while AI applications showed promising accuracy for basic cancer information, the tools struggled to address complex scenarios.
They also warned that most of the studies were early-phase investigations, reflecting the rapidly evolving nature of the field and the difficulty for published research to keep pace.
While current research showed preclinical feasibility and acceptability of AI tools for patients, the review found limited data on effectiveness.
Education/information
The researchers came up with mixed evidence when reviewing the ability of large language models to generate reliable patient-facing cancer information.
They referred to studies that showed 88% of ChatGPT’s recommendations for breast cancer screening and prevention were considered appropriate, with the chatbot producing comparable accuracy for lung cancer prevention guidance.
However for cervical cancer screening information, the researchers found despite performing well with delivering basic educational content, AI applications were less effective when asked about complex clinical scenarios and nuanced cases.
Symptom checking and triage
Evidence showed certain advanced AI programs could offer highly accurate recommendations on supportive care and management of treatment-related symptoms, suggesting a potential use for routine symptom monitoring.
Another promising application was the integration of AI with electronic patient-reported outcomes, which appeared an effective way to enhance patient-physician communication while reducing the documentation burden, the researchers said.
Telehealth and virtual care
The researchers said while telehealth adoption in oncology care accelerated during the COVID-19 pandemic, there was now strong provider and patient interest in an expansion of virtual care to symptom monitoring and communication support.
Research showed some virtual care systems had particular promise in providing regular patient check-ins, conversation transcription, and mental health support.
Research and clinical trials
Clinical trial accessibility and participation was another critical domain in cancer care where large language models showed potential, the researchers said, giving the example of making the complex process of trial matching more efficient.
The researchers also pointed to a clinical pilot that demonstrated GPT-generated trial summaries, after human clinical review, could enhance patients’ comprehension of a trial, which could be used as a tool for consent processes.
Other uses
The review noted many other applications of AI were being evaluated, including personalised summaries of medical appointments in plain language that could address the challenge of information overload for patients and their caregivers.
Other examples included AI systems navigating insurance claims, coordinating appointments across multiple specialists, and managing medication schedules, which the researchers said could reduce the administrative burden and improve care continuity – although there would likely be technical and regulatory hurdles to overcome.
The researchers said evidence on emotional support tools was also encouraging but required careful safety and ethical consideration, particularly without human oversight.
Ethical considerations
One of the major concerns of the reviewers was privacy.
They said for AI applications to generate contextually appropriate responses, these systems required access to sensitive patient information, which raised questions over whether robust protection measures were in place as well as around the level of patients’ awareness on how their data were being stored, used and protected.
“Of particular concern is the growing use of publicly available large language models (such as ChatGPT) by patients and caregivers seeking medical information—these platforms… may store and repurpose entered information for model training, and offer no guarantees of data privacy or security,” the authors said.
Other ethical concerns outlined in the review included the ability of AI to hallucinate – that is to produce convincing but factually incorrect information, as well as inequitable access which risked a two-tiered system where AI tools were only available to those with reliable internet and the necessary infrastructure.
The researchers have also put together a practical discussion checklist to help guide patient-provider conversations about the use of AI applications in cancer care. Topics include:
- Current AI usage and purpose
- Comfort level with AI accuracy and potential need for clinician verification
- Nature of questions being asked of AI
- Sharing AI-generated insights with the healthcare team
- Caregiver involvement using AI tools
- Privacy and data handling
- Satisfaction and engagement