What an AI discharge plan can accomplish when a clinician remains in the room

October 29, 2026 PubMed publication record · Single-center randomized trial; 90 post-surgical patients · Primary source

After stomach or colorectal cancer surgery, nutritional guidance can become another burden for someone who is already recovering. A randomized trial tested a clinician-reviewed, AI-facilitated discharge plan that translated individual needs into quantified food and protein targets. The comparison was standard generic advice—not equally individualized counseling without AI.

Ninety people were randomized. At 30 days, 66.7% of the intervention group met at least 90% of their energy targets compared with 40.0% of controls (odds ratio 3.0, 95% CI 1.3–7.1). Protein adequacy was 62.2% versus 35.6% (OR 2.9, 95% CI 1.2–6.8). Clinician time per patient was 8.2 versus 12.5 minutes. Patient-reported fatigue, appetite and digestive symptoms also favored the intervention, while weight change, blood nutritional markers and complications did not differ significantly.

The exact contribution of the language model cannot be isolated: the intervention combined individualized attention, explicit quantitative goals, generated text and clinician review. That is not a weakness to ignore; it is the central implementation question.

Clinical reading: Interesting supervised discharge-support research, not evidence that unsupervised chatbot diet advice improves recovery.