Tuning into flavor: predicting coffee sensory attributes from EEG with boosted-tree regression models

Flavor, a multimodal perception based on taste, smell, and tactile cues, plays a significant role in consumer preferences and purchase intentions toward coffee. In this exploratory study, we assessed the potential of electroencephalography (EEG) and machine learning (ML) techniques to predict coffee sensory attributes. We extracted spectral and temporal features from a professional panel while tasting coffee samples and basic water solutions. We trained multiple Least-Squares Boosted Trees (LSBoost) and optimized their hyperparameters through a 100-step Bayesian approach based on a Leave-One-Subject-Out (LOSO) scheme. The models achieved, overall, high predictive accuracy (MAE < 0.75 on a 0 − 10 scale) and medium-to-large robustness (Cohen's d >0.6) with respect to mean and lasso benchmark regressors. Feature importance analysis revealed that spectral powers and Hjorth's parameters within parietal, central, and frontal regions were the most predictive. Our findings endorse the use of EEG-based ML models as an alternative to traditional flavor evaluation methods, such as Descriptive Sensory Analysis (DSA).