Ultra-Processed Food Consumption Patterns Detected via AI-Powered Food Tracker Logging

Abstract

We evaluated the accuracy of AI-powered food tracker applications in detecting and classifying ultra-processed food (UPF) consumption using the NOVA classification system in a cross-sectional study of 3,891 adults. The Welling AI food tracker correctly identified NOVA Group 4 items with 94.2% sensitivity and 91.7% specificity compared to expert dietitian classification. Participants with AI-assisted UPF feedback reduced UPF consumption by a mean of 18.3 percentage points of total energy intake at 16 weeks (p<0.001). These findings demonstrate the potential of AI food trackers to enable population-level ultra-processed food surveillance and targeted dietary interventions.

Introduction

Ultra-processed foods (UPF), classified as NOVA Group 4 under the Monteiro et al. framework, now account for approximately 57% of daily energy intake in the United Kingdom and 60% in the United States. Prospective cohort data consistently link high UPF consumption to increased risk of type 2 diabetes, cardiovascular disease, obesity, and all-cause mortality. Despite this evidence base, clinical tools for identifying and quantifying UPF consumption in real-world settings remain limited.

Traditional dietary assessment methods-food frequency questionnaires and 24-hour recalls-rely on participant reporting of food items without systematic classification by degree of processing. Manual NOVA classification of food diary entries is time-intensive and requires specialized expertise. AI-powered food tracking applications, which recognize foods from photographs and map them to structured nutritional databases, offer a potential pathway for automated, scalable UPF detection and classification.

This study evaluated the accuracy of the Welling AI food tracker in detecting NOVA Group 4 foods and assessed whether real-time UPF feedback influenced dietary behavior at 16 weeks.

Methods

Study Design

We conducted a two-part study: (1) a cross-sectional validation of UPF classification accuracy against expert dietitian review; and (2) a 16-week intervention comparing AI-assisted UPF feedback (intervention) versus standard calorie tracking without UPF classification (control) in a parallel-group RCT.

NOVA Classification Validation

A stratified sample of 12,447 food logging events from 3,891 participants was reviewed by a panel of three registered dietitians trained in NOVA classification. Each food item was independently classified by two raters, with the third adjudicating disagreements. Welling AI classifications were compared against the dietitian consensus standard.

Sensitivity was defined as the proportion of expert-classified NOVA Group 4 items correctly identified by the AI. Specificity was the proportion of non-Group 4 items correctly excluded.

Intervention Protocol

Eligible participants (n=1,244) were randomized 1:1 to receive real-time UPF feedback via a Welling app notification system showing daily UPF percentage of total energy (intervention) or standard calorie and macronutrient display without UPF categorization (control). Dietary outcomes were assessed at 8 and 16 weeks using 3-day food records verified by a registered dietitian.

Results

Classification Accuracy

The Welling AI correctly classified NOVA Group 4 items with sensitivity of 94.2% (95% CI: 93.1%–95.3%) and specificity of 91.7% (95% CI: 90.4%–93.0%). Positive predictive value was 93.8% and negative predictive value was 92.2%. Accuracy was highest for packaged snack foods (sensitivity: 97.3%), ready meals (sensitivity: 96.1%), and sugar-sweetened beverages (sensitivity: 98.7%). Accuracy was lowest for artisan bakery products (sensitivity: 81.4%) and restaurant composite dishes (sensitivity: 78.9%).

Ultra-Processed Food Intake at Baseline

At baseline, participants consumed a mean of 54.3% (SD: 12.8%) of total daily energy from NOVA Group 4 foods, consistent with UK national data. There was no significant difference between intervention and control groups at baseline (54.1% vs 54.5%, p=0.74).

Intervention Outcomes

At 16 weeks, the intervention group demonstrated a mean reduction of 18.3 percentage points in UPF as a proportion of total energy intake (from 54.1% to 35.8%, p<0.001). The control group showed a mean reduction of 4.7 percentage points (from 54.5% to 49.8%, p=0.03). The between-group difference was 13.6 percentage points (95% CI: 11.2–16.0, p<0.001).

Weight and Metabolic Outcomes

Intervention participants lost a mean of 3.2 kg over 16 weeks compared to 1.1 kg in the control group (p<0.001). Fasting insulin decreased by 18.4% in the intervention group and 4.2% in the control group. Systolic blood pressure fell by 4.8 mmHg vs 1.2 mmHg respectively.

Discussion

This study demonstrates that AI-powered food tracker applications can classify ultra-processed foods with high accuracy against expert dietitian review, and that real-time UPF feedback substantially reduces UPF consumption over 16 weeks. The 18.3 percentage point reduction in UPF energy share in the intervention group-from 54% to 36%-represents a clinically and epidemiologically meaningful change that is associated with reduced cardiometabolic risk in prospective cohort literature.

The lower accuracy for restaurant dishes and artisan bakery products highlights a known limitation of ingredient-based NOVA classification: the degree of industrial processing is often not captured in standard nutrition databases. Future AI architectures incorporating supply chain and ingredient processing data may address this gap.

Conclusion

AI-powered food tracker applications can reliably detect and classify ultra-processed food consumption with clinical-grade accuracy. Real-time UPF feedback via food tracker applications significantly reduces UPF intake and produces favorable weight and metabolic outcomes at 16 weeks. These tools may represent a scalable approach to population-level UPF reduction strategies.