Can you calculate nutritional content of meals using ChatGPT? 🍽️

Better or worse than human error.

Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images

Study Details

This new study evaluated the performance of 3 LLMs…

1️⃣ ChatGPT-4o

2️⃣ Claude 3.5 Sonnet

3️⃣ Gemini 1.5 Pro

…in estimating nutritional content including…

⚖️ Food weight

🔥 Energy content

🔬 Macronutrient composition

…from standardised food photographs 📸

🥦 52 total photos were used (16 x food components, 36 x complete meals) 🥘

Each estimation was compared to direct weighing and nutritional database analysis 🌟

Here is what they found ⬇️

Key Findings

Both ChatGPT and Claude demonstrated similar accuracy (mean absolute percentage error) values of…

⚖️ ~37% for weight estimation

🔥 ~36% for energy estimation

Gemini showed substantially higher errors across all nutrients (MAPE 64.2%-109.9%) ⚠️

Correlations between model estimates and reference values ranged from…

👉 0.65 to 0.81 for ChatGPT and Claude

👉 0.58-0.73 for Gemini

ALL models exhibited systematic underestimation that increased with portion size 🚨

…with bias slopes ranging from -0.23 to -0.50.

Conclusion

ChatGPT and Claude achieved accuracy levels comparable with traditional self-reported dietary assessment methods 📝

BUT without associated user burden, suggesting potential use as dietary monitoring tools 🔍

This may be useful in future but accuracy would need to improve before being used in practice‼️

If we want to provide the highest quality of feedback and nutritional guidance, we need high quality data ✅

Reference

https://pubmed.ncbi.nlm.nih.gov/41081011/

---

Performance Nutrition Digest is powered by Healthspan Elite ⚡️

Healthspan Elite is a science-led sports nutrition range developed in collaboration with leading experts to support athletic performance, recovery, and health.

Use code: TCDIGEST25 for 25% off at https://www.healthspanelite.co.uk/ 💪

---