AI-guided nutrition helped more patients hit intake targets—but was AI the reason?
In a 90-patient randomized trial after gastrointestinal cancer surgery, a multicomponent AI-guided programme improved 30-day energy and protein intake. It did not isolate AI from individualization, targets or clinician review.
The short version
- What happened
- Ninety patients were randomized 1:1 after laparoscopic gastric or colorectal cancer surgery, and all completed the 30-day follow-up.
- Who it affects
- Patients received more specific guidance after major cancer surgery, but the study does not show that a model can safely replace a dietitian or that better 30-day intake improves longer-term recovery.
- What remains unproven
- Single centre
Keep in mind
Single centre
90 patients
30-day follow-up
Multicomponent intervention
AI effect not isolated
Participants and delivery staff unblinded
Not powered for clinical complications
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
More patients reached the short-term intake target
A randomized trial in China found that more patients recovering from gastrointestinal cancer surgery reached prescribed energy and protein targets after receiving an AI-facilitated nutrition plan. Of 90 adults undergoing laparoscopic gastric or colorectal cancer surgery, 45 received the programme and 45 received the hospital's standard generic discharge advice. All completed the 30-day follow-up. Energy adequacy—the prespecified primary outcome, defined as at least 90% of a target of 25–30 kilocalories per kilogram per day—was reached by 66.7% in the intervention group and 40.0% in standard care. The odds ratio was 3.0, with a 95% confidence interval from 1.3 to 7.1.
Protein adequacy, defined as at least 1.2 grams per kilogram per day, was reached by 62.2% and 35.6% respectively, an odds ratio of 2.9 with a 95% confidence interval from 1.2 to 6.8. Blinded dietitians calculated intake from three-day food records during the fourth week after discharge. Patients in the intervention group also reported less fatigue, nausea or vomiting, appetite loss and gastrointestinal symptoms. Provider time averaged 8.2 minutes, compared with 12.5 minutes for standard advice.[1]
The comparison tested a package, not an AI ingredient
The intervention was not simply a chatbot answering questions. A nurse entered surgical details, laboratory values, diet tolerance and cultural preferences into an Analysis–Blueprint–Concrete Plan prompt for DeepSeek v3.2. The output included quantified energy, protein and fluid targets, phased written advice and a clinician safety review before delivery. Standard care was a one-page generic handout with brief verbal guidance and no routine individualization or quantified targets.
That contrast makes the result useful as a test of the combined service, but weak evidence that AI was the active cause. Individualized counselling, written targets, longer guidance or the structured review could each have contributed. Participants and staff delivering the intervention were not blinded, although dietitians assessing intake, laboratory staff and the statistician were. The study was powered for dietary adequacy, not complications, and lasted only 30 days.[1]
The distinction that matters
What changed at 30 days
The randomized comparison measured target achievement, not long-term recovery.
Energy
66.7% vs 40.0% reached at least 90% of target
27-point difference
Protein
62.2% vs 35.6% reached 1.2 g/kg/day
27-point difference
Clinical outcomes
No significant weight, serum-marker or complication difference
Longer-term benefit unproven
Better intake did not yet mean better recovery
There were no statistically significant differences in weight change, serum albumin, haemoglobin or postoperative complications. Weight loss averaged 2.1% in the intervention group and 2.5% in standard care. Complications occurred in roughly one fifth of each group. Those results do not prove equivalence; with 90 patients, uncommon harms or modest clinical benefits could easily be missed. Readmissions were also too uncommon for a dependable comparison.
For patients and clinical teams, the practical promise is a quicker way to assemble specific, reviewable discharge guidance. It is not permission to let a model prescribe nutrition without oversight. Confidence would rise with a larger multicentre trial that compares AI-assisted guidance with equally individualized, quantified human counselling, follows weight and clinical outcomes for longer, and measures errors, staff workload and patient understanding. Until then, the trial supports the package's short-term effectiveness, not an independent AI advantage or improved recovery.[1]