INSIGHTS FROM MACHINE-LEARNED DIET SUCCESS PREDICTION.

Area of Science:

• Nutrition Science
• Data Science
• Behavioral Science

Background:

• The rise of fitness apps and the quantified self movement has led to increased data logging for health and weight management.
• Publicly shared food diary data offers a rich resource for understanding dietary patterns and behaviors.

Purpose of the Study:

• To analyze public food diary data to identify characteristics of successful and unsuccessful diets.
• To train a machine learning model to predict adherence to self-set calorie goals based on logged food intake.
• To uncover features within food diaries that correlate with consistently exceeding or falling short of calorie targets.

Main Methods:

• Utilized a dataset of public food diaries from over 4,000 long-term active MyFitnessPal users.
• Developed and trained a machine learning model to predict repeated over/under consumption relative to daily calorie goals.
• Identified and analyzed key features contributing to the model's predictions, including food types and app usage patterns.

Main Results:

• Identified specific food tokens (e.g., "mcdonalds") and categories (e.g., "dessert") associated with exceeding calorie goals.
• Discovered less obvious dietary predictors, such as the differential impact of pork versus poultry consumption on dieting success.
• Found that using the "quick added calories" feature is indicative of exceeding calorie targets.

Conclusions:

• Public food diary data can be effectively mined to understand dieting behaviors and outcomes.
• Machine learning models can predict diet success based on logged food intake and user behaviors.
• Findings provide novel insights into factors influencing weight management and suggest potential areas for personalized dietary guidance.