Large language model-assisted chemotherapy side effect management in solid tumor patients: The ALLIANCE randomized controlled trial

ALLIANCE

Doente / População Intervenção / Exposição Comparação Desfecho

In adults with breast or colorectal cancer starting chemotherapy, clinician access to ChatGPT Plus for side effect management did not improve EORTC QLQ-C30 global health status compared with care without LLM access.

N

60doentes

Desenho

Single-centre, 1:1 randomised two-stage trial with patients but not clinicians blinded, in adults with breast or colorectal cancer starting ≥3 months of chemotherapy (ECOG 0-2), comparing clinicians who could consult ChatGPT Plus with de-identified symptom data at visits against clinicians without that option, reported as a pre-specified 60-patient interim analysis of safety and feasibility after which enrolment stopped (NCT07198581).

Desfecho

Change in EORTC QLQ-C30 global health status (GHS) from baseline to end of treatment.

ResultadoThe primary endpoint was not met. The adjusted GHS difference at end of treatment was -18.4 (95% CI -31.2 to -5.5; n=31, 55%). The estimate went against the LLM arm and no patient-level benefit was shown. At mid-treatment the adjusted GHS difference was -5.3 (95% CI -18.2 to 7.7; n=48). Of 60 enrolled, 56 were analysed, 28 per arm. Recorded adverse events were 150 vs 78 (5.4 vs 2.8 per patient; P = 0.006). In a post hoc classification there were 48 vs 27 distinct terms for toxicities with no routine management, against 29 vs 22 for routine toxicities. Dose reduction occurred in 12/28 vs 14/28. The LLM was used at 47 of 99 monthly assessments (47.5%) and 7 of 28 patients never used it. Advice was adopted in full at 8 of 47 uses and partly at 29. Output was appropriate at 34 of 47 and never inappropriate. Hallucination was absent at 44 and mild at 3.

Kang Y, Song JM, Park JY, et al. Presented at ESMO Congress 2026, abstract LBA19, 24 October. Not yet published.

Discussão e crítica

This small single-centre interim analysis of 56 patients was designed to assess safety and feasibility, not efficacy, and on its primary endpoint it showed no benefit: quality of life at end of treatment was worse in the LLM arm. The end-of-treatment estimate rests on only 31 patients (55%), so missing data and early stopping make it fragile in either direction. Clinicians were not blinded, so the near doubling of recorded adverse events may reflect closer recording rather than more toxicity, and the split into toxicities with and without a routine was post hoc. The abstract gives no outcomes for the specific toxicities the LLM is said to have helped with, so the claim that it adds value where clinicians have no routine remains a hypothesis to test.