Comparing ChatGPT and DeepSeek for ultra-processed food classification: AI models for nutritional research and...

There is growing evidence linking the consumption of ultra-processed foods (UPFs) to adverse health outcomes. Accurate classification of foods according to the extent and purpose of industrial processing is therefore essential for improving dietary assessment and public health strategies. This study aimed to evaluate and compare the performance of two large language models (LLMs), DeepSeek-R1 and ChatGPT o1, in classifying foods according to the NOVA classification system.

Both LLMs were tasked with categorizing a standardized list of 1,168 food items obtained from the Brazilian Food Composition Table (TBCA, version 7.0). The classifications generated by the models were compared with a reference list manually classified by a trained researcher. Quantitative analyses included the calculation of unweighted Cohen's kappa between the LLMs, as well as accuracy, sensitivity, specificity, precision, and F1 score for each model. Qualitative analyses were conducted to explore discrepancies in food classification.

ChatGPT o1 demonstrated superior performance across all evaluated metrics, achieving an accuracy of 98.0%, sensitivity of 94.7%, specificity of 99.0%, and an F1 score of 95.6%. In comparison, DeepSeek-R1 achieved an accuracy of 92.6%, sensitivity of 69.8%, specificity of 99.3%, and an F1 score of 81.1%. ChatGPT o1 also produced substantially fewer misclassifications than DeepSeek-R1 (23 versus 86, respectively).

The findings highlight the potential of large language models to support dietary assessment and nutrition research. The development of an automated tool based on the NOVA food classification framework is recommended to assist nutritionists and researchers, enabling faster and more consistent food classification in both clinical and research settings.