(P05-007-23) Accelerating Classification of Food Processing Levels Using a Fine-Tuned Language Model: A Multi-Country Study

Objectives: Consumption and availability of ultra-processed foods (UPFs) have increased in most countries. High consumption of UPFs has been associated with noncommunicable diseases. However, the food processing level determination process relies on expertise-based manual categorization, which is labor-intensive and time-consuming, delaying research on its implications on health outcomes. This study aims to utilize a fine-tuning transformer-based language model to automate the classification of food processing levels according to the NOVA system.

Methods: Food product information from FLIP-Canada 2013-2020 (n=119,541), FLIP-Latin America and the Caribbean (FLIP-LAC) 2018-2022 (n=8,466), and USDA Branded Food Products Database (USDA-BFPD) 2017-2021 (n=1,671,501) databases were used. The levels of processing of packaged foods and beverages were manually coded according to the NOVA food classification system, then used to train and test the model. In the fine-tuned language model, we used the ingredients text information found on food labels for the modified pre-trained BERT model and added a classification layer on top, and then fine-tuned the whole model end-to-end on the given datasets. To evaluate the performance of algorithms on the given task, accuracy, balanced accuracy, F1 score, confusion matrix, and receiver operating characteristic curves were calculated.

Results: The fine-tuned language model achieved the highest F1 score of 0.98 in predicting food processing levels of FLIP-Canada 2017, using manually-coded NOVA as inputs. This outperformed the structured nutrient facts data model, bag-of-words model, and pre-trained language model. The fine-tuned language model reached a higher generalizable ability than traditional models on the external test sub-datasets with manually-coded NOVA (F1 scores: 0.94 vs 0.89 for FLIP-Canada 2020 >50% subset, 0.89 vs 0.69 for FLIP-LAC >50% subset, and 0.80 vs 0.70 for USDA-BFPD >0.1% subset).

Conclusions: This study indicated that the fine-tuned language model achieved high accuracy in the NOVA classification of food processing levels, using ingredient information found on food labels. This automation strategy is effective and generalizable, which can accelerate the monitoring and regulation of UPFs in the global food supply.

Funding Sources: Canadian Institutes of Health Research (CIHR)