Coffee stands as the major component of the global hot drink market, with a worldwide production exceeding 176 million bags and consumption reaching 7 billion kilograms (Kim Y. et al., 2025). Among over 60 coffee plant species, only 10 are extensively cultivated, with Coffea arabica (arabica), Coffea canephora (robusta), and Coffea liberica being the most prevalent (Feria-Morales, 2002). Arabica and robusta make up 99% of global production (Jayakumar et al., 2017) and commercial coffee typically results from blending their beans in varying proportions (Seninde and Chambers, 2020).

Consumer preference and purchasing intentions for coffee are mainly driven by subjective factors, such as taste, aroma, and body (Li et al., 2019). These elements belong to the broader concept of flavor, a multimodal experience in which gustatory, olfactory, trigeminal, and somatosensory inputs are individually processed before being integrated (Small, 2012). Gustatory signals ascend via the nucleus of the solitary tract (NST) and ventroposteromedial nucleus (VPM) to the primary taste cortex in the rostral insula and adjoining frontal operculum, where the identity and intensity of basic tastes, as well as oral texture and temperature, are encoded. Retronasal olfactory inputs, initially processed in the piriform cortex, converge with gustatory signals in the orbitofrontal cortex (OFC), which also integrates oral somatosensory and trigeminal inputs such as viscosity, temperature, irritation, and astringency (Rolls, 2005). Within the OFC, convergent inputs give rise to multimodal flavor representations, in which sensory modalities are integrated and assigned hedonic value. At the same time, projections to the amygdala and anterior cingulate cortex (ACC) further embed these representations within affective and motivational systems (Small, 2012).

Flavor assessment often relies on Descriptive Sensory Analysis (DSA), wherein expert panels assign numerical scores to standardized sensory attributes (Yang and Lee, 2019). Various coffee attributes have been suggested (Spencer et al., 2016). However, only bitter, sour, sweet, and astringent (i.e., mouth-drying sensation) have established reference solutions (Batali et al., 2022).

Based on self-reports, DSA can be confounded by physiological and psychological biases. Physiological phenomena include sensory adaptation and multimodal enhancement/suppression. Psychological phenomena include expectation, stimulus/proximity/logical errors, habituation, halo effect, presentation order, mutual suggestion, and central/extreme rating tendency (Stone et al., 2021; Civille et al., 2024). To mitigate these risks, international standards (e.g., ISO 13299, ISO 11132, and ISO 8586) and the scientific literature recommend extraneous cues blinding, randomized or Williams-balanced presentations, adequate rests/rinses alternation, and ongoing panel performance monitoring (Sipos et al., 2021). Additionally, direct techniques based on bioelectrical measures have recently been advocated (Torrico et al., 2023; Rodrigues et al., 2024).

Previous studies have explored the use of electroencephalography (EEG) for flavor assessment. Global field power and scalp topographies (Crouzet et al., 2015), as well as phase in the delta band (Wallroth et al., 2018) and spectral powers in alpha and theta bands (Yang et al., 2023) have emerged as candidate neurometrics. Similar results, involving alpha, beta, and theta powers, have been observed in coffee tasting tasks (Hsu and Chen, 2021; Tonacci et al., 2024). However, their correlational—rather than causal—nature poses a risk of reverse inference problems (Poldrack, 2006). Pattern-decoding methods based on Machine Learning (ML) models have been suggested to mitigate this issue (Nathan and Del Pinal, 2017). Furthermore, being free of rigid theoretical assumptions, ML methods could also be helpful in revealing latent structures in the data, providing new theoretical insights and hypotheses (Verzelli et al., 2024).

Research on flavor prediction with ML and EEG data is still limited (see the following Section 2 for details). Most of the studies employed basic water solutions as eliciting stimuli, and the few examining coffee focused on other target variables than taste. Moreover, nearly all existing models were classifiers to discriminate among basic tastes (e.g., sour, sweet, bitter, salty, umami, and neutral) instead of predicting the intensity level of sensory attributes. Therefore, such models are ill-suited to replace or even complement traditional DSA.

To address these limitations, we performed an exploratory study recording the EEG data from expert coffee tasters while they tasted both reference solutions and coffee samples. We trained multiple tree-based ensemble regressors to predict the intensity level of bitter, sweet, acid and astringent, achieving high performances and robustness against benchmark models. We interpreted the fitted models, identifying as most informative, spectral and temporal features within parietal, central and frontal regions.

As previously mentioned, most of the past studies on taste prediction using EEG data and ML methods trained classifiers to discriminate among basic water solutions. De et al. (2023) fed temporal (maximum/minimum values, mean, kurtosis and skewness) and spectral [Power Spectral Densities (PSDs) in theta, delta, alpha and beta bands] features into a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) to discriminate sour, sweet, bitter, salty, umami, and neutral solutions from 46 subjects. They obtained an accuracy of 97.16%. Xia et al. (2024) employed a Convolutional Neural Network (CNN) with spatiotemporally augmented raw EEG data to identify sour, sweet, bitter, and salty solutions from 20 subjects. They reached 99.5% of accuracy. Li et al. (2025) trained a Support Vector Machine (SVM) with spectral features (wavelet decompositions in α and θ bands) to classify sour, sweet, bitter, salty, and umami solutions from 22 subjects. They reported a maximum accuracy of 76.13%. Vo et al. (2023) trained a feed-forward Neural Network (NN) using spectral features (powers in delta, theta, alpha, beta, and gamma bands) to discriminate between salty and sour solutions from 15 people. The accuracy was 84.36%.

Only one study moved from discrimination to intensity level prediction. Zhao et al. (2022) contrasted linear, tree, and ensemble regressors, trained with temporal and information-related features (energy, absolute mean value, and wavelet entropy), to predict the intensity level of sour, sweet, bitter, salty, and umami from 10 subjects. The best model, Extreme Gradient Boosting (XGBoost), achieved a goodness-of-fit (measured through the R² coefficient), ranging from −0.22 to 0.18.

Two studies focused on other-than-basic water solutions. González-España et al. (2023) aimed to discriminate wine vs. water and wine vs. wine tasting tasks of 10 participants through an SVM with temporal and spatial features (global field powers and channel averages). They reported accuracies greater than the chance level of 70% for both predictions. Naser and Aydemir (2024) trained a k-Nearest Neighbors (kNN) and a Random Forest (RF) with temporal and spectral features (mean value of the Hilbert-transformed EEG signal and level-2 wavelet coefficients) to discriminate four food substances (oils of Orange, Mint, Thyme, and Clove) from 10 subjects. The highest accuracy, obtained with the kNN, was 87.5%.

Coffee was selected as an eliciting stimulus in two studies. However, as previously mentioned, the target belonged to other aspects than taste. Maram et al. (2023) trained a CNN with raw EEG data to classify the preference of 3 coffee brands from 12 participants, obtaining an accuracy of 83.43%. Xu et al. (2021) compared several Bayesian Regression (BR) models, trained with spectral features (powers in theta, alpha, beta, and gamma bands), to predict the emotional responses to tasting tasks from 32 subjects. The best model achieved a goodness-of-fit [measured through the Watanabe-Akaike Information Criterion (WAIC)] of 963.55.

Four models reported significant improvements from the benchmarks, with an overall MAE of 0.537 ± 0.073 (anti-log:0.714 ± 0.124) and d scores of d_Me = 0.858 ± 0.341, d_LR = 0.897 ± 0.326. The targets were Bi (trained on Be), Sw (on Be), Ac (on Co), and As (on BeCo).

The Ac prediction obtained the best MAE (Co: 0.459 ± 0.178), while Sw the worst one (Be: 0.600 ± 0.282). The highest robustness against benchmark regressors was achieved by Bi trained on Be (d_Me = 1.344, d_LR = 1.372), while the lowest one was observed for Sw trained on Be (d_Me = 0.651, d_LR = 0.660).

The following Table 1 reports the best hyperparameters and the training time (in seconds) of the significant models, split for dataset and target. The following Table 2 summarizes the performances (cross-validated MAEs and Cohen's d coefficients) of the significant models, split for datasets and targets.

Both spectral and temporal features contributed to the predictions, but their importance scores varied substantially across datasets and targets. The highest median score was observed for p_α (Med = 0.139, IQR = 0.190), whereas the lowest was for M (Med = 0.034, IQR = 0.037). Age and Group appeared as predictors in two models each: Age in Sw with Be and Ac with Co, while Group in Ac with Co and As with BeCo. Group achieved the highest importance, not only compared to Age but also across all features (Med = 0.252, IQR = 0.136). Table 3 summarizes the channel-averaged feature importance scores, split by dataset and target.

The topographic plot in Figure 3 qualitatively identified central, occipital, parietal, and frontal regions as most important for the overall prediction.

In this study, we trained Least-Squares Boosted Trees (LSBoost) with spectral and temporal EEG features to predict sensory attributes—bitterness (Bi), sweetness (Sw), acidity (Ac), and astringency (As)—of Coffee (Co) and basic solutions (Be). The best configuration of hyperparameters and data normalization was obtained through a Bayesian optimisation approach, following a Leave-One-Subject-Out (LOSO) scheme. The LSBoost's performances were compared with mean (Me) and lasso (LR) regressors through Cohen's d coefficients, and the feature importance for type and channel was assessed from the trees' coefficients.

The significant models achieved high performances, with an average anti-log MAE of about 7% of the scale range and a medium-to-large (d_{Me, LR}>0.5) (Cohen, 1992) robustness against the benchmarks. The lowest MAE of Ac is in line with previous studies that identified sour as the best predictable flavor, with the highest R² coefficient in Zhao et al. (2022) and the second (after salty) highest accuracy in Li et al. (2025). Compared to other dimensions, Sw has already shown poor performances (De et al., 2023) and low feature discriminability (Xia et al., 2024), supporting the obtained highest MAE and lowest d coefficients. Finally, the robustness of Bi against benchmarks may reflect the well-known evolutionary adaptation in vertebrates toward heightened bitter taste sensitivity for early toxin detection and avoidance (Wooding et al., 2021).

The feature importance of p_α, p_β, and p_γ is in line with past studies that effectively trained ML models using spectral powers in α, β, and γ bands (De et al., 2023; Vo et al., 2023). The role of p_δ, and p_θ as key predictors is supported by previous research studies that identified differences in δ and θ bands during flavor evaluation (Wallroth et al., 2018; Yang et al., 2023). Overall, the involvement of EEG features from specific central, parietal, and frontal regions has already been observed in predictive (Li et al., 2025) and experimental (Lejap et al., 2024) studies. The contribution of temporal parameters, represented by C, M and A, matches the good performance of past deep-learning models (e.g., CNNs and RNNs) trained with the raw EEG signal (De et al., 2023; Xia et al., 2024). Finally, the significance of Group and Age could be related to the previously reported influence of expertise (Croijmans and Majid, 2016) and aging (Mojet, 2003) on sensory evaluations. However, the reasons why they impacted only in two models require further investigation.

This study acknowledges some limitations. First, despite being in line with DSA studies that typically involve 5–15 experts (Gacula and Rutenbeck, 2006), the sample size must still be considered limited. Increasing it in both magnitude and heterogeneity (e.g., adding non-expert tasters and accounting for their coffee-consumption frequency) would potentially improve not only the performance but also the generalisability of the models. An increase in dataset size would also yield less noisy results, which is particularly relevant in chemo-sensory studies. In fact, despite the use of advanced denoising techniques and the selection of short epochs with minimal muscular artifacts, the data quality in the present study should still be regarded as suboptimal. Second, the experiment has not accounted for confounders given by the fixed Be-Co order, the temperature difference between the Be and Co samples, as well as potential visual cues (e.g., the colors of the liquids), potentially biasing the sensory analyses (Delwiche, 2023). Future confirmatory studies should be, thus, based on a fully-randomized and truly-blind design. Third, our models were trained and validated in a single session per subject. Although being standard practice in multivariate-pattern-analysis with brain-imaging data (Taxali et al., 2021), it prevented us from quantifying the test-retest reliability of the models. Future works should acquire longitudinal recordings—at least a second session separated by days or weeks—to determine the stability of features and models over time.

Nevertheless, our exploratory study endorses the use of regression techniques based on EEG data in flavor assessment, as an alternative to self-report sensory evaluations.