Interpretable AI for Behavioral Prediction: An Ethical Laboratory Experiment on Snack Choice Prediction

Introduction: The application of artificial intelligence (AI) in behavioral prediction has shown promise across domains like mental health, autonomous vehicles, and consumer behavior. However, challenges such as algorithmic bias, lack of interpretability, and ethical concerns persist. This study addresses these gaps by developing an interpretable AI model to predict snack choices in a controlled laboratory experiment. Methods: A random forest classifier was trained to predict participants’ snack choices (healthy vs. unhealthy) based on contextual factors (hunger, mood, time of day) and historical choices. Data were collected from 75 adults over 10 sessions, with features engineered to capture both immediate and longitudinal patterns. Model performance was evaluated using accuracy, precision, recall, and feature importance analysis. Results: The model achieved 85.33% accuracy, with hunger level, historical choices, and mood identified as the most influential predictors. Performance improved over sessions (peaking at 93.33% accuracy in sessions 8–9), highlighting the value of longitudinal data. Subgroup analyses showed consistent performance across age, gender, and BMI, with higher accuracy for participants with healthier habits and higher socioeconomic status. Conclusions: This study demonstrates the feasibility of interpretable AI models in predicting dietary behavior while addressing ethical concerns through rigorous data anonymization and informed consent protocols. The findings underscore the potential of AI to inform personalized interventions for healthier eating habits and provide a framework for ethical AI implementation in behavioral research.