Artificial Intelligence Falls Short on Complex Nutrition Guidance

Study for Man and MachineArtificial intelligence is hard at work, reshaping healthcare delivery, as well as the delivery of profits from healthcare. So what does this mean for the delivery of dietary services? Can Chat GPT and other artificial intelligence technologies help with the complex task of delivering nutrition guidance for health?

Maybe. Or maybe not yet. People have been trying to apply artificial intelligence and machine learning to nutrition guidance for some time now. But the results so far are underwhelming. In a new study of generative AI for dietary advice, researchers conclude:

“Although ChatGPT exhibited a reasonable accuracy in providing general dietary advice for NCDs, its efficacy decreased in complex situations necessitating customized strategies.”

Single Conditions
and Complex Cases

This study by Valentina Ponzo and colleagues was quite straightforward. They evaluated the recommendations that ChatGPT, version 3.5, generated in response to each of seven different clinical conditions: dyslipidemia, hypertension, type 2 diabetes, obesity, NAFLD, chronic kidney disease, and sarcopenia. For each condition, they prompted the system with three different patient queries about dietary recommendations:

“Could you provide guidance on planning an optimal diet to manage…?”
“What are the dietary recommendations for …?”
“I have …, what should I eat?”

Researchers also prompted the system to give advice for someone with a combination of issues:

“I have type 2 diabetes mellitus, obesity, and chronic kidney disease. Can you give me nutritional advice?”

Two dietitians and a doctor with nutrition expertise evaluated the consistency of the output from ChatGPT with current clinical practice guidelines. ChatGPT did well enough on recommendations for specific conditions – only contradicting clinical guidelines once for obesity and once for NAFLD. But in the case of a patient with multiple conditions, the advice was contradictory or inappropriate.

A Work in Progress

The authors note that generative AI is moving on. ChatGPT-4 incorporates learning from more recent information and has more flexibility for adaptation to specific tasks. This technology will continue to evolve.

But these findings, as well as efforts to use machine learning to develop “precision nutrition” recommendations for health, serve as a reminder that dietary guidance presents a complex challenge. Repeatedly we find that hubris and hype can lead us to unintended consequences.

So we’ll be taking the power of artificial intelligence in this field with a grain of salt.

Not to exceed guidance for sodium intake, of course.

Click here for the study and here for more on the challenges of precision nutrition.

Study for Man and Machine, painting by Hannah Hoch / WikiArt

Subscribe by email to follow the accumulating evidence and observations that shape our view of health, obesity, and policy.


 

February 13, 2024