"EVALUATING THE ACCURACY OF THE AI-BASED APPLICATION VS HUMAN INTERPRETATION IN THE FOOD PORTION SIZE ESTIMATION"

Abstract

Background Accurate portion size estimation is crucial for dietary assessment, particularly in nutrition tracking and chronic disease prevention. Traditional methods that rely on human estimations often suffer from significant biases, resulting in inaccurate data. With the advent of AI-based mobile applications like CaloScanAI, the potential to enhance dietary assessment through automated tools has grown. However, the accuracy of these tools compared to human estimations and how demographic factors such as gender and age impact performance remains uncertain and under-explored. Methods Pictures of fifty-four food items and fifteen drink items commonly consumed in Kazakhstan were captured using RGB (red, green, blue) cameras and incorporated into questionnaires. Sixty-four participants estimated portion sizes for solid, liquid, and mixed food items, while an app (CaloScanAI) performed the same task via image recognition. Estimation accuracy was evaluated using accuracy metrics, including Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE), which were computed for the overall accuracy between humans and the app for each food category. Additionally, Mean Relative Error (MRE%) was calculated to analyze over- and underestimation patterns. A series of statistical tests, including Wilcoxon signed-rank tests (for paired samples), Mann-Whitney U tests (for independent two-group comparisons), and Kruskal-Wallis tests (for three or more groups’ differences), were employed to assess the results and performance disparities between humans and AI. Statistical significance was evaluated at p < 0.05. 3 Results The overall accuracy of CaloScanAI estimations was higher than that of human estimations in terms of absolute error (MAE: 69.02g vs. 132.24g) and in relative terms (MAPE: 51.73% vs. 102.42%) while also producing fewer large errors (RMSE: 76.20% vs. 170.50%). The findings reveal systematic biases in human portion estimation, with a consistent tendency toward overestimation, particularly for smaller portions (e.g., Toast: 66.8g MAE, 607.6% MAPE). AI estimations were generally more stable but exhibited misclassification for complex and mixed dishes. Males tended to overestimate portion sizes more completely (e.g., Toast: male 1042% MAPE and female 366.3%), and females were more accurate in certain food items. AI also had higher accuracy over the standardised food items and poorer accuracy with amorphous and liquid-based foods, like Herring Salad (-42.4% MRE). Conclusion This research demonstrates the potential of AI in dietary assessment while also highlighting current limitations in AI and human estimation methods. Findings suggest that multi-angle imaging or depth-sensing technologies could improve AI accuracy. Future research should focus on integrating AI-based tools with real-time dietary tracking systems to provide more personalized and accurate dietary assessments.