Related Experiment Video
Updated: Jun 12, 2026
Estimate the Cognitive Load Using Electrocardiographic Measure: A Human-AI Collaborative Task
Published on: December 5, 2025
Evaluating a Multitask Artificial Intelligence Model Compared With Humans for Portion-Size Estimation
Bibinur Nurmanova1, Zhuldyz Omarova1, Aibota Sanatbyek2
1Department of Biomedical Sciences, School of Medicine, Nazarbayev University, Astana, Kazakhstan.
Background:
Accurate dietary assessment is essential for precision nutrition and nutrition surveillance. Portion-size estimation remains challenging, particularly in Central Asia, where communal eating and nonstandard household measures are common. Advances in artificial intelligence (AI), especially multitask learning models capable of simultaneous food recognition and portion estimation, offer promising alternatives to traditional self-report methods. However, direct comparisons between AI and human estimation remain limited.
Objectives:
This study compared 3 methods: unassisted human judgment, visual food atlas assistance, and an AI model, using Central Asian food items.
Methods:
A total of 128 participants from Astana, Kazakhstan, visually estimated portion sizes of 51 foods and 8 beverages from standardized photographs. Participants were randomly divided to unassisted or atlas-assisted estimation, and an AI model trained on Central Asian food images was evaluated. Actual food weights served as the reference. Accuracy was assessed using mean absolute error (MAE) and mean absolute percentage error (MAPE) across food types, categories, and portion sizes.
Results:
Atlas-assisted estimation showed the highest overall accuracy (MAE = 80.81 g; MAPE = 44.76%), whereas unassisted estimation was least accurate (MAE = 133.86 g; MAPE = 79.40%). The AI model performed intermediately (MAE = 97.37 g; MAPE = 67.81%). Differences across methods were significant (P < 0.05): atlas-assisted estimation was consistent, whereas AI model accuracy varied by food category and portion size. The AI model performed best for beverages (MAE = 34.48 g; MAPE = 11.79%) and medium portions, with higher errors for small portions and visually complex foods. Differences between females and males were significant in the unassisted group (P = 0.038), with males reporting larger portions; differences were minimal with atlas assistance (P = 0.114).
Conclusions:
Visual atlases use substantially improved portion-size estimation. The AI model requires refinement for complex foods and small servings. Integrating visual and AI-based tools can enhance region-specific dietary monitoring strategies.