O028 - CAN ARTIFICIAL INTELLIGENCE SAFELY PRESCRIBE PARENTERAL NUTRITION? A COMPARISON WITH EXPERT CLINICAL PRACTICE

Linked sessions

O028

CAN ARTIFICIAL INTELLIGENCE SAFELY PRESCRIBE PARENTERAL NUTRITION? A COMPARISON WITH EXPERT CLINICAL PRACTICE

S. U. CELIK1,*, B. Kutlu2, G. Oral3, P. Arı4, S. Demirer1

1General Surgery, Ankara University School of Medicine, 2General Surgery, Acıbadem Hospital, 3General Surgery, Istanbul university-Cerrahpasa, 4Clinical Nutrition, Ankara University School of Medicine, Ankara, Türkiye

 

Rationale: Parenteral nutrition (PN) is a complex, high-risk therapy requiring individualized assessment. Although artificial intelligence (AI) may support clinical nutrition, its reliability in PN prescribing remains unclear. This study compared PN prescriptions generated by ChatGPT-5.4 with those prepared by an expert clinical nutritionist.

Methods: Hospitalized patients requiring PN were included. Basal metabolic rate was calculated using the Schofield equation, and daily target calories were determined accordingly. Prescriptions generated by the clinical nutritionist and ChatGPT-5.4 were assessed independently and blinded to each other. To evaluate consistency, the same patient data were entered into ChatGPT-5.4 three times, and agreement among repeated outputs was analyzed. Comparisons focused on first-day calorie, protein, carbohydrate, and lipid recommendations.

Results: Seventy-five patients were included (mean age 60.3 ± 13.9 years; 64% male). According to Subjective Global Assessment, 6.7% were SGA-A, 13.3% SGA-B, and 80.0% SGA-C. Mean albumin was 3.0 ± 0.4 g/dL, mean weight loss over the previous 3–6 months was 16.71 ± 12.26%, mean basal metabolic rate was 1363.37 ± 196.29 kcal/day, and mean target calorie requirement was 1771.60 ± 262.50 kcal/day. ChatGPT-5.4 showed poor agreement for first-day calories (ICC=0.487), protein (ICC=0.340), and carbohydrates (ICC=0.373), and moderate agreement for lipids (ICC=0.617). Passing–Bablok regression showed proportional bias for calories (slope 0.56, intercept 449.58), protein (0.38, 58.22), carbohydrates (0.47, 14.93), and lipids (0.57, 29.26).

Image:



 

Conclusion: ChatGPT-5.4 did not match expert clinical judgment in complex PN prescribing and should currently be considered a supportive tool rather than a replacement for specialist nutritional evaluation.

Disclosure of Interest: None declared