Patients recovering from surgery for gastric or colorectal cancer often leave the hospital with little more than a standardized pamphlet about eating small, frequent meals and choosing soft foods. A randomized controlled trial published in Supportive Care in Cancer tested whether a more tailored approach, drafted by a large language model and checked by a clinician, could do better. In 90 patients, the AI-facilitated strategy significantly increased the proportion of people meeting energy and protein intake targets at 30 days after discharge and improved several patient-reported symptoms, while cutting the time clinicians spent producing the guidance.
The trial was conducted at Tongji Hospital in China and enrolled patients scheduled for elective laparoscopic gastrointestinal cancer surgery between December 2025 and February 2026. Of 187 patients screened, 90 met the eligibility criteria and were randomly assigned in a 1:1 ratio to either the AI-guided group or a standard care group. Randomization was stratified by surgery type, allocation was concealed in sequentially numbered opaque sealed envelopes, and outcome assessors, including the dietitians who analyzed dietary records and the statistician running the primary analysis, remained blinded to group assignment. Participants and the staff delivering the interventions could not be blinded, given the obvious difference between a printed personalized document and a generic handout. All 90 randomized patients completed the 30-day follow-up, with no losses.
The intervention was deliberately multicomponent. Within 24 hours before discharge, a research nurse collected each patient’s age, sex, weight, surgery type and extent, current diet tolerance, laboratory values such as hemoglobin, albumin and prealbumin, and any special considerations including diabetes or dumping syndrome risk. These data were entered into a structured prompt template built on a three-module framework, Analysis, Blueprint, Concrete Plan, abbreviated ABC, which had been refined through pilot testing with 10 patients. The prompt was submitted to DeepSeek v3.2, a Chinese-language-optimized large language model with a mixture-of-experts architecture totaling 671 billion parameters, of which 37 billion are activated per token. A clinical dietitian then reviewed each AI-generated output for safety, a process taking roughly three to five minutes per patient, focusing on identifying dangerous recommendations rather than rewriting content. The final document was printed, handed to the patient, and explained verbally by the nurse.
The control group received the institution’s conventional discharge advice: a standardized one-page handout of about 310 words describing general postoperative nutrition principles, delivered verbally by the attending surgeon or ward nurse over five to fifteen minutes. No individualization based on patient-specific factors occurred, and quantified energy or protein targets were not routinely provided. The contrast, the authors emphasize, was therefore not AI versus human expertise, but a combined AI-plus-individualization-plus-quantification strategy versus generic advice.
The primary outcome was energy intake adequacy at 30 days post-discharge, defined as achieving at least 90 percent of a target of 25 to 30 kilocalories per kilogram of body weight per day, calculated from discharge weight in line with ESPEN guidelines. Participants kept three-day dietary records during the fourth week after discharge, covering two weekdays and one weekend day, and blinded dietitians analyzed the records with nutrition analysis software. In the AI-guided group, 66.7 percent of patients reached the energy target versus 40.0 percent in standard care, an odds ratio of 3.0 with a 95 percent confidence interval of 1.3 to 7.1 and a P value of 0.01. Protein intake adequacy, defined as consuming at least 1.2 grams of protein per kilogram per day, was also higher in the AI-guided group, at 62.2 percent versus 35.6 percent, an odds ratio of 2.9 with a confidence interval of 1.2 to 6.8 and P equal to 0.01.
Patient-reported outcomes at 30 days also favored the AI-guided group. On the EORTC QLQ-C30 questionnaire, the intervention group reported significantly lower symptom scores for fatigue, nausea and vomiting, and appetite loss. On the Gastrointestinal Symptom Rating Scale, the AI-guided group had a lower total symptom score, 2.1 versus 2.5 on the seven-point scale, with significant improvements in the diarrhea and indigestion subscales. Global quality of life trended higher in the AI-guided group but did not reach statistical significance, and differences in the disease-specific modules for gastric and colorectal cancer patients were not significant. The authors note that the roughly half-standard-deviation reduction in gastrointestinal symptom burden may be clinically meaningful, but that these benefits were observed over only 30 days.
Notably, several hard clinical measures showed no differences. Weight change from discharge to 30 days was similar between groups, at roughly minus 2.1 percent versus minus 2.5 percent, and changes in serum albumin and hemoglobin did not differ significantly. Postoperative complications within 30 days occurred at comparable rates, about 20 percent in the AI-guided group versus 22 percent in standard care, and nutrition-related readmissions were likewise similar. The trial was powered for dietary intake adequacy and followed patients for only one month, so these clinical comparisons are exploratory and cannot demonstrate benefit or harm.
Process measures revealed where the intervention differed most. Two independent blinded dietitians rated the AI-generated documents as significantly more complete, more individualized, and more actionable than the standard handouts, while accuracy and safety scores did not differ between groups. Quantified nutritional targets appeared in 93.3 percent of AI-guided documents versus 51.1 percent of standard care guidance. Provider time per patient fell from 12.5 minutes to 8.2 minutes, a reduction of about one-third, even though the AI-generated documents were far longer, averaging 845 words versus 310. Prespecified subgroup analysis found consistent effects across gastric and colorectal cancer patients, with no significant interaction by surgery type, and sensitivity analyses including a per-protocol analysis did not materially change the results.
The authors are explicit about the trial’s central limitation: it cannot isolate the independent effect of AI. The intervention bundled AI generation, individualization with patient-specific clinical and cultural data, explicit quantified targets, phased written guidance, and clinician review, while the control was generic advice. The observed benefits may reflect individualization, goal setting, the sheer quantity of information, additional patient attention, or some combination, rather than AI itself. The study also does not show that AI-generated guidance outperforms similarly individualized counseling delivered by a human dietitian without AI. Other limitations include the single-center design, a study population with relatively favorable baseline nutritional status, a 30-day follow-up that cannot capture weight recovery, micronutrient status, or late complications, dependence on self-reported dietary records, and the use of a single AI model whose performance may not generalize to other systems or to non-Chinese dietary contexts.
No serious adverse events related to the nutritional guidance interventions occurred, and an independent data safety monitoring board confirmed no safety concerns. The authors frame AI-assisted guidance as a practical delivery tool that may extend the reach of limited dietitian expertise rather than replace it, and they call for factorial trials that dismantle the intervention into its components, multicenter studies in diverse settings, longer follow-up extending to six or twelve months, comparisons across different large language models, and formal cost-effectiveness analyses. Within those boundaries, the trial establishes that a structured, individualized, quantified discharge strategy facilitated by AI improved short-term dietary target achievement and reduced symptom burden, while clinical supervision remained essential throughout.
Subject of Research: AI-facilitated individualized nutritional guidance after gastrointestinal cancer surgery
Article Title: AI-facilitated individualized nutritional guidance with ABC-structured prompts improves short-term dietary intake adequacy and patient-reported symptoms after gastrointestinal cancer surgery: a randomized controlled trial
Article References: Chen, H., Jiang, W., Li, M., & Zhang, S. (2026). AI-facilitated individualized nutritional guidance with ABC-structured prompts improves short-term dietary intake adequacy and patient-reported symptoms after gastrointestinal cancer surgery: a randomized controlled trial. Supportive Care in Cancer, 34(11), Article 1170. https://doi.org/10.1007/s00520-026-11401-7
Image Credits: AI Generated
DOI: 10.1007/s00520-026-11401-7
Keywords: artificial intelligence, nutritional guidance, gastrointestinal cancer surgery, randomized controlled trial, dietary intake, patient-reported outcomes, DeepSeek, discharge planning, energy intake adequacy, protein intake, large language models, postoperative nutrition
News Source: Daisy Hatcher. (October 29, 2026). AI-Crafted Discharge Nutrition Plans Boost Short-Term Dietary Intake After Gastrointestinal Cancer Surgery. Scienmag.
