Large language models (LLMs) used by clinicians to identify adverse events in patients with cancer receiving chemotherapy appeared safe when used under physician supervision, according to interim findings from the single-center, randomized controlled ALLIANCE trial presented at the ESMO Congress 2026 (Abstract LBA19). Physicians rated the LLM’s output as appropriate or moderately appropriate in all cases, and no adverse events were attributed to the use of the LLM.  

However, the trial did not demonstrate that the use of the LLM provided clinical benefit, as chemotherapy-related outcomes remained largely unchanged and patient quality of life did not improve.  

Background

Clinicians are increasingly using LLMs in patient care. One recent study of U.S. clinicians’ use of OpenEvidence indicated that 952,000 U.S. clinicians used the platform over a 30-month period. Oncology and hematology clinicians showed higher relative use compared with other specialties.  

A total of 4,609 studies have examined clinical LLMs, but only 19 of these were randomized trials.  

Although LLMs are evidently widely used by oncology clinicians, no randomized trial has yet explored clinician-facing use in oncology, said presenting author Young-Joon Kang, MD, PhD, of the Division of Breast Surgery, Department of Surgery, Incheon St. Mary’s Hospital and the College of Medicine, The Catholic University of Korea. Until now.  

The single-center, randomized controlled ALLIANCE trial evaluated clinicians’ use of an LLM to identify chemotherapy-related adverse events in patients with solid tumors, with the goal of improving adverse event management and quality of life. 

Dr. Kang noted that in an analysis of three clinical trials, 41% to 74% of patient-reported toxicities were not recorded by their physicians. He also noted that when symptoms are detected and acted upon, quality of life improves and patients can remain on chemotherapy longer.  

Study Methods and Safety Considerations

The investigator-initiated ALLIANCE trial enrolled adult patients with breast or colon cancer who were scheduled for at least 3 months of chemotherapy and had no cancer-related symptoms at baseline. Participants were randomly assigned 1:1 to receive either LLM-assisted care or standard care. In the intervention arm, physicians could consult ChatGPT Plus during visits as needed. Randomization was stratified by cancer type. Patients were blinded to their study arm, and all were required to complete EORTC QLQ-C30 questionnaires assessing global health status at baseline, mid-treatment, and the end of treatment.  

Five treating physicians participated in the study, and 174 patients were planned for enrollment, with 60 required for the interim analysis. The primary endpoint was the change in global health status scores from baseline to the end of treatment, and secondary endpoints included adverse events, dose modifications and intensity, emergency visits and admissions, LLM use, output ratings, and physician satisfaction.  

Minimal rules were established for the use of ChatGPT Plus, and the participating physicians were not required to use structured prompts. All chats had to be based on de-identified information only, and chats were not saved to the chat history or connected to patients’ electronic health records. Physicians reviewed every output before taking any action to assess its appropriateness and check for hallucinations; all decisions remained with the treating physician.  

LLM Use Findings

Four patients were excluded after randomization, leaving 56 patients in the interim analysis. The study arms were largely similar in terms of patient characteristics, except that there were more female patients in the LLM-assisted care arm.  

ChatGPT Plus was used in fewer than half (47%) of the monthly reports in the LLM-assisted care arm. Of the 28 patients in this arm, 21 had at least one instance of LLM use for their case. Across 99 total monthly physician reports, the LLM was not used in 52, most commonly because it was “not needed.” A technical problem was the only other reported reason for nonuse, occurring once.  

Each LLM session took a median of 3 minutes, and overall LLM use did not decline over the study period.  

“When it was opened, the output was safe on review,” Dr. Kang said. LLM output was rated as appropriate in 72% of monthly reports in which the LLM was used, and no output was considered inappropriate. There were three recorded “mild” hallucinations but no cases of “moderate or severe” hallucinations. Physicians adopted the LLM’s advice in full in 17% of monthly reports, partially in 62%, and did not adopt it in the remainder.  

No adverse events were attributed to the intervention, although these ratings were not independently reviewed.  

Patient Outcome Findings

More toxicities were identified in the LLM-assisted care arm (mean = 5.4 vs 2.8; Mann-Whitney P = .006). After adjustment for sex and cancer type, the rate ratio was 1.50 (95% confidence interval [CI] = 0.99–2.27; interaction P = .65).  

“More toxicity was recorded, and the increase was larger where no routine exists,” Dr. Kang said. More of the adverse events recorded in the LLM-assisted care arm were those without an established management routine (ie, antiemetics for nausea or granulocyte colony-stimulating factor for neutropenia). In the LLM-assisted arm, 29 identified adverse events had an existing routine, whereas 48 did not; in comparison, 22 events identified in the standard care arm had an existing routine, whereas 27 did not.  

Forty-three percent of patients in the LLM-assisted care arm underwent a dose reduction compared with 50% in the standard care arm (Fisher’s exact P = .79).  

Changes in global health status from baseline did not favor the LLM-assisted arm at either time point. At the mid-treatment assessment (n = 48), the difference in global health status scores was –5.3 (95% CI = –18.2 to 7.7). At the end of treatment assessment (n = 31), the difference was –18.4 (95% CI = –31.2 to –5.5). Twenty-five patients were still receiving chemotherapy at the end of the study period.  

Limitations, Takeaways, and Next Steps

Although the LLM helped clinicians identify more toxicities, its use did not improve toxicity management or patient quality of life.  

“Put simply, on-request LLM support for the clinician was not useful in real-world everyday practice,” Dr. Kang commented.  

Following the prespecified interim analysis, trial enrollment was stopped because the study did not demonstrate a positive change in global health status.  

Because of the study design, physicians could choose whether to use ChatGPT Plus, and the LLM was used in fewer than half of the monthly assessments. “We wanted a trial to look like everyday practice, so we did not mandate use and did not prescribe the prompt,” Dr. Kang commented. “The price is a diluted intervention.” 

The study was also limited by its single-center design, small sample size, and lack of physician blinding, “so we cannot tell whether more toxicity was found or simply more was written down,” Dr. Kang explained.  

Participating physicians also noted that the AI tool needs to be easily accessible in a busy clinic, preferably through integration with the electronic health record. However, Dr. Kang noted that such integration is constrained by data protection requirements and ethics reviews in many regions.  

Going forward, the researchers plan to assess patient-led LLM use between visits, with clinicians remaining in the loop, and to reevaluate the safety of clinician-facing LLM use in a setting where patients significantly outnumber specialists.  

Insights

“I would really like to compliment the prospective and randomized design of this study, even though the number of patients is small—final 56 patients—but they are really trying to show how these tools that are already being used behave in our clinical workflows,” commented invited discussant Vanja Miskovic, PhD, Assistant Professor at Politecnico di Milano and co-leader of the AI-ON-Lab at IRCCS Istituto Nazionale Tumori, regarding the ALLIANCE trial. “Aside from that, I think the biggest strength of this abstract is a very ambitious patient-centered endpoint.” 

Dr. Miskovic asked whether the participating physicians were trained to use AI tools and whether their clinical and AI expertise influenced their decisions to use or not use the LLM. She questioned whether the study was assessing who uses LLMs, when, and why. “I think this can open a lot of new research around this topic,” she said.  

She also emphasized Dr. Kang’s observation that increased recording of toxicities did not necessarily indicate better detection but might simply reflect more extensive documentation of adverse events.  

Additionally, she questioned the robustness of the unfavorable quality-of-life estimate, given that only 55% of patients were assessed at the end of treatment, raising the possibility of bias.  

“I think that we, in the future, need to target the right clinical problem. We need to standardize the AI intervention, in my opinion. So maybe the LLM arm should be more standardized, controlled prompts, exposure, and also educate clinicians on how to use these LLMs. Measure meaningful outcomes, like it was done. And finally, validate the clinical impact, compare with standard of care, and assess harms a bit more in depth,” Dr. Miskovic recommended. 

DISCLOSURES: Funding for the trial came from the Translational R&D Project through the Institute for Bio-Medical Convergence, Incheon St. Mary’s Hospital, The Catholic University of Korea, and from the National Research Foundation of Korea. Dr. Kang reported financial relationships, not related to this trial, with Ataraxis AI, the Korean Society of Surgical Oncology (KSSO), and the Korean Association of Robotic Surgeons (KAROS).  

Dr. Miskovic has been an invited speaker for Novartis, Merck, and Daiichi Sankyo, and is a principal investigator for a local Pfizer grant. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.