← Back to Home

AI Increases Documentation of Chemotherapy Side Effects, but Are Patients Better Off? Randomized Trial Finds No Care Benefit

An interim analysis of 56 patients in the ALLIANCE trial found that allowing physicians to consult ChatGPT increased documented chemotherapy toxicities but did not improve management or quality of life. This small study highlights a gap between filling in medical records and improving patients’ circumstances that still requires evidence to bridge.

By SURL BioNews

During chemotherapy, identifying one more side effect may be the starting point for adjusting care; without subsequent action, however, patients may not feel any better. The ALLIANCE randomized trial, presented at the 2026 European Society for Medical Oncology (ESMO) Congress, illustrated this gap: allowing physicians to consult a large language model (LLM) as needed increased documentation of chemotherapy toxicities but showed no improvement in patients’ quality of life or toxicity management.

This single-center study in South Korea enrolled adult patients with breast or colon cancer who were preparing to receive at least three months of chemotherapy. Patients were assigned in a one-to-one ratio to an LLM-assisted group or a standard-care group and were treated by the same five physicians. The interim analysis excluded four patients after randomization, leaving 56 patients. The primary endpoint was the change in global health status from before treatment to the end of treatment, measured using the EORTC QLQ-C30 questionnaire.

The study deliberately preserved the flexibility of routine outpatient practice. In an interview with VJOncology, researcher Young-Joon Kang explained that physicians decided for themselves whether to consult ChatGPT, and prompts were not standardized. Of 99 monthly reports in the assisted group, only 47 documented use of the tool; the reason for not using it was almost always that physicians considered it unnecessary. Assignment to the assisted group therefore did not mean that every assessment involved AI assistance.

The results showed an average of 5.4 documented toxicities per patient in the assisted group, compared with 2.8 in the standard-care group. Most of the additional toxicities lacked established management protocols. However, no improvement was seen in dose adjustments, and ESMO’s account of the same analysis also noted that toxicity management and quality of life did not improve as a result. The study cannot yet distinguish whether the increased documentation reflected the detection of more symptoms or more complete recording.

The quality-of-life results require particular caution. According to ASCO AI in Oncology, the difference in global health scores at the end of treatment even favored the standard-care group. At that point, however, only 31 patients had completed the assessment, while another 25 were still receiving chemotherapy, making the estimates harder to interpret because of incomplete data. After the prespecified interim analysis, the research team stopped enrollment because no improvement in global health status was observed; this is insufficient to establish that the LLM itself worsens quality of life.

The safety findings also apply under specific conditions: physicians could enter only de-identified data, conversations were not connected to electronic medical records, and outputs had to be reviewed before any action was taken. The study recorded three mild hallucinations, no moderate or severe hallucinations, and no adverse events attributed to the intervention. However, these assessments were made by the participating physicians without independent review and cannot support an inference that unsupervised use would be equally safe.

ALLIANCE provides a specific and limited answer: under this model, in which physicians decided when to use the tool, providing access to an LLM has not yet translated into patient benefit. The small sample, use in fewer than half of reports, and physicians’ awareness of group assignments all limit the conclusions. The next step is to clarify when the tool should intervene and how identified symptoms can be linked to effective management, then validate its value through patients’ actual experiences and clinical outcomes.

References

  1. ASCO AI in Oncology
  2. European Society for Medical Oncology (ESMO)
  3. VJOncology