EEG feature extraction and machine learning for consumer preference analysis of whisky

Abstract

Understanding consumer preference for food and beverages is central to product development, yet subjective self-reports alone often lack objectivity. Electroencephalography (EEG) can reveal implicit neural signals of preference. However, existing flavor studies utilizing EEG are mostly confined to group difference analysis, cannot identify stable preference-related EEG markers or achieve objective preference prediction. To address this gap, this study integrated sensory evaluation, systematic EEG feature extraction and machine learning to quantify and predict consumer whisky preference. Twenty-four panelists were tested with 12 whisky samples via 9-point hedonic and 5-point JAR scales. Time-domain and frequency-domain features were extracted from EEG recordings, and four classification models were trained to predict preference. Compared with the accuracy of models built on raw EEG data (≈51%), feature extraction significantly boosted performance, with the Random Forest model achieving the highest test accuracy of 82.6%. Feature weight analysis further identified potentially predictive EEG indicators, including frontal α asymmetry and α band power. This work verifies the predictive capacity of EEG features for machine learning modeling, thereby establishing a neurophysiologically based method for the objective evaluation. This scheme complements traditional subjective sensory assessment and delivers actionable technical guidance for industrial flavor tuning and product development. However, the high homogeneity of participants in age and background limits the generalizability of our experimental results.

Graphical abstract

Highlights

• •EEG combined with machine learning enables objective prediction of consumer preference for whisky.
• •Extracted time-domain and frequency-domain features greatly outperform raw EEG data in classification tasks.
• •The Random Forest model reaches 82.6% test accuracy; frontal α asymmetry and α power are key neural markers.
• •This study offers a novel approach for objective sensory evaluation in the beverage industry.

1. Introduction

Consumer preference is the core guiding principle for product development and market positioning in food flavor, and its data can directly guide the regulation of flavor compounds and process optimization, thereby enhancing product market competitiveness and consumer acceptability (Raz et al., 2008). Traditional sensory evaluation methods mainly rely on subjective scoring and preference ranking (Ismail et al., 2001; Khan and Shafiur Rahman, 2021; Sipos et al., 2025). Although these approaches are widely applied, they are easily affected by personal experience, emotion, testing environments and inconsistent evaluation criteria, thus restricting their objectivity and reliability (Trumbo et al., 2021). In contrast, neural signals reflect perceptual processes, serving as objective and functionally meaningful indicators (Pashkovski et al., 2020). Therefore, the application of objective physiological indicators to achieve accurate characterization and discrimination of sensory preference has become an important research direction in the field of food and beverage quality evaluation (Modesti et al., 2022; Zhao et al., 2024).

Currently, numerous studies have incorporated biometric technologies into traditional sensory evaluation, including electroencephalography (EEG) (Wang et al., 2025a), functional near-infrared spectroscopy (fNIRS) (Mai et al., 2025), physiological signals (Stuldreher et al., 2024), and facial expression analysis (Gonzalez Viejo et al., 2019). These technologies have advanced sensory research from descriptive analysis and preference assessment to direct detection of intrinsic neurophysiological mechanisms, thereby enabling multi-dimensional evaluation of sensory preference. Among these technologies, EEG has emerged as an important tool for the objective evaluation of emotion, cognitive processing, and sensory preference due to its high temporal resolution, non-invasiveness, and real-time data acquisition capabilities, which allow for the dynamic capture of neural oscillatory responses induced by sensory stimuli (Songsamoe et al., 2019). Iannilli et al. (2015) demonstrated divergent brain activation patterns triggered by food and non-food odors. Yavuz and Aydemir, (2016) extracted EEG features using continuous wavelet transform and applied k-NN to discriminate brain responses to different food odors.

Existing EEG studies in the flavor-related field have demonstrated that α-waves activity increases with increasing alcohol content, and compared with ethanol solution of the same concentration, Baijiu with complex flavors elicits more significant brain responses (Wang et al., 2025a). Additionally, individual aroma compounds in Baijiu, such as ethyl caprylate and ethyl acetate, as well as their mixtures, can induce significant changes in EEG microstate parameters (Wang et al., 2025b). Furthermore, brain regions including the prefrontal cortex and parietal cortex also exhibit significant activation during flavor integration and preference judgment (Small and Prescott, 2005). On the other hand, due to substantial individual differences in neural responses, EEG is also used to investigate differences in food and beverage perception among distinct populations. For instance, coffee aroma stimulation elicits distinct brain activity patterns between genders: males primarily exhibit changes in fast α-waves, while females show alterations in low β-waves (Hong et al., 2025); When consuming beer, beer experts display more efficient neural activity during taste processing, and the correlation between their sensory scores and the neural responses associated with pleasure processing and memory, as reflected by EEG, is significantly higher than that of ordinary consumers (Hinojosa-Aguayo et al., 2022). However, existing EEG flavor perception research mainly centers on Baijiu and beer, while literature targeting whisky consumer preference is extremely scarce. Only a small number of studies explored neural responses to general alcoholic stimuli via EEG microstate analysis (Wang et al., 2025c), with no dedicated work predicting whisky flavor preference.

Despite the promising potential of EEG in sensory preference identification, current research still has limitations: although current flavor-related studies have conducted time-domain and frequency-domain analyses of EEG, most are limited to difference tests and do not involve further EEG feature extraction. This hinders the exploration of stable and reliable electrophysiological markers for sensory preference and prevents the objective quantification and prediction of individual preference based on EEG signals. Whisky possesses a complex flavor system shaped by diverse raw material origins and oak maturation, and exhibits distinct alcohol concentration and volatile components relative to Baijiu and beer, while relevant EEG-based investigations remain limited. Accordingly, this study aims to explore whether consumers exhibit distinct EEG patterns under different whisky preference states, whether EEG features are significantly associated with subjective sensory scores, and whether consumer preference can be objectively predicted by combining EEG signals with machine learning algorithms. To this end, this study comprehensively integrates sensory evaluation, EEG feature extraction, and multiple classification models to analyze neural responses elicited by typical whisky samples. Ultimately, this work intends to establish a neurophysiological analytical framework for beverage preference evaluation, providing objective technical support for flavor perception research and industrial product optimization.

2. Materials and methods

2.1. Subjects

A total of 24 consumers were recruited as participants in this study, all of whom were students at Jiangnan University aged 20-24 years, including 12 females and 12 males. This recruitment strategy was adopted to minimize confounding variations related to age and sensory experience. Nevertheless, this relatively homogeneous cohort imposes limitations on generalizability to broader consumer groups of varying ages and backgrounds. All participants were right-handed, had no history of alcohol allergy, mental illness, neurological disorders, respiratory diseases, or cardiovascular diseases, and had not taken any drugs affecting brain function. Meanwhile, they exhibited normal olfactory and gustatory abilities and were capable of completing sensory evaluation. Individuals with heavy drinking or smoking habits in the recent period were excluded. The sensory analysis protocol of this study was approved by the Ethics Committee of Jiangnan University (JNU202506RB010), and all procedures strictly adhered to the ethical guidelines for human experimentation outlined in the World Medical Association's Declaration of Helsinki. Several precautions were adopted to reduce cumulative alcohol intake, including mandatory expectoration and adequate resting intervals. Participants could withdraw freely at any time. Recruited undergraduates came from different departments, and all procedures complied with the approved ethical protocol.

To reduce the interference of food intake and time factors on experimental results, all EEG tests were conducted from 10:00 to 11:00 and 15:00 to 16:00 (Wang et al., 2025a). Participants were required to finish their meals 2-3 h before the test and avoid spicy and irritating foods. Before the experiment, all participants provided informed consent, completed a basic information questionnaire, and received detailed experimental guidance to fully familiarize themselves with the evaluation scales, experimental procedures and estimated duration. During the test, the experimental environment was controlled to eliminate external disturbances such as noise and visual stimulation (Noble et al., 2024). Participants were instructed to follow experimental requirements strictly, keep their bodies still, and focus on perception. After the experiment, all participants received corresponding remuneration for their involvement.

2.2. Samples

Twelve representative whisky samples were included in the study (Table 1), covering four producing countries, three style categories and diverse cask types to ensure sample diversity and representativeness. The alcohol concentration of all whisky samples ranges from 40% to 45% ABV. Minor differences in ethanol content across samples may introduce small confounding effects on EEG recordings.

Table 1.

Whisky samples selected for sensory evaluation and EEG recording.

Country	ID	Brand	Product	Type	Cask Type
USA	1	Jim Beam	Jim Beam White Label	Blended	American Oak Cask
	2	Jack Daniel's	Jack Daniel's Old No.7	Blended	American Oak Cask
	3	Maker's Mark	Maker's Mark	Bourbon	American Oak Cask
UK	4	Glenlivet	The Glenlivet Founder's Reserve	Single Malt	American Oak Cask
	5	Macallan	The Macallan 12 Year Old Sherry Oak	Single Malt	Sherry Cask
	6	Laphroaig	Laphroaig 10 Year Old	Single Malt	Bourbon Cask
Japan	7	Yamazaki	Yamazaki 1923	Single Malt	Bourbon Cask + Sherry Cask + Red Wine Cask
	8	Hibiki	Hibiki Japanese Harmony	Blended	Bourbon Cask + Sherry Cask + Mizunara Oak Cask
	9	Hakushu	Hakushu 1973	Single Malt	Bourbon Cask + Sherry Cask
Ireland	10	Jameson	Jameson Original	Blended	Bourbon Cask + Sherry Cask
	11	Tullamore D.E.W.	Tullamore D.E.W. Original	Blended	Bourbon Cask + Sherry Cask
	12	Bushmills	Bushmills 10 Year Old	Single Malt	Bourbon Cask + Sherry Cask

2.3. Sensory evaluation

The experiments were conducted in a standardized sensory laboratory at a temperature of 25 ± 1°C. All samples were stored at room temperature to ensure a consistent testing temperature. Sensory evaluation was conducted synchronously with EEG recording. After tasting and expectorating each sample, participants immediately completed rating scales, then rinsed their mouths with purified water before the resting interval. A 9-point hedonic scale (1 = dislike extremely, 5 = neither like nor dislike, and 9 = like extremely) and a 5-point just-about-right (JAR) scale (1 = much less than I like, 3 = just about right, 5 = much more than I like) were adopted for sensory assessment. The evaluated attributes included pungency liking, aroma liking, taste liking, overall liking, pungency intensity, aroma intensity, and taste intensity. The complete sensory questionnaire is provided in Supplementary Fig. A1. Intraclass correlation coefficient (ICC) was calculated to evaluate inter-rater reliability of sensory ratings.

2.4. EEG recording

Before the experiment, all participants washed their hair with fragrance-free shampoo to ensure the scalp was dry and oil-free. This procedure reduced electrode contact impedance and guaranteed the quality of EEG signals. Subsequently, a 32-channel EEG cap was fitted for each participant. EEG signals were collected via the iRecorder W32 wireless EEG/ERP recording system (Niantong Intelligent Technology Co., Ltd., Shanghai, China), with supporting eConScan W32 software applied for signal recording, impedance monitoring and event marking. Electrode placement followed the international 10–20 system (Klem et al., 1999), and the sampling rate was set at 500 Hz. Device calibration was conducted after EEG cap placement to make sure all electrode impedance was kept below 5 kΩ. Data recording was initiated once qualified signal recording conditions were met.

Participants tasted 12 samples in a random order, with each sample tested in triplicate, and were kept unaware of specific sample information throughout the trial. The single-trial procedure was implemented as follows: 1. Resting phase (30 s): Participants rinsed their mouths with ultrapure water and remained in a quiet resting state to restore EEG signals to the baseline level. 2. Tasting and EEG recording phase (15 s): A start marker was inserted into the EEG recording system. A 10-s resting baseline was recorded first, during which participants stayed calm and relaxed. After 10 s, 5 mL of the sample was accurately delivered into the oral cavity through a liquid delivery device for tasting. The 5-s tasting epoch was selected according to previous neuro-sensory studies on alcoholic beverages, and this duration was further validated as reasonable in our preliminary experiments. Participants were required to remain still with a fixed head position to avoid motion artifacts, and neural electrical activity was continuously recorded synchronously. After tasting, participants spit out the sample and rinsed their mouths again with ultrapure water to eliminate residual flavor substances. 3. Subjective evaluation and recovery phase: Participants completed hedonic and JAR questionnaires based on their sensory perceptions while staying relaxed. The next sample stimulation started only after EEG signals returned to a stable baseline. The above procedures were repeated until all sample tests were completed. The detailed experimental design is illustrated in Fig. 1. To eliminate variations in EEG signals induced by different testing days, all 36 trials were completed within a single experimental session. A minimum 60-s resting interval was set between adjacent trials to mitigate fatigue, and participants were permitted extra rest time upon request.

Fig. 1.

Overview of the EEG experimental procedure.

2.5. EEG data preprocessing

Raw EEG data were preprocessed using EEGLAB 2021.0 toolbox in MATLAB R2020b. The preprocessing pipeline was implemented as follows. First, raw data were imported, and the international 10-05 electrode template was matched to define channel locations. Next, 0.2-40 Hz band-pass filtering was applied. Epochs were extracted based on event markers. The resting baseline captured prior to each tasting trial was used for baseline correction of the corresponding tasting-related EEG signals to mitigate progressive signal drift over the experiment. Epochs contaminated by artifacts were manually rejected, with approximately 15% of all epochs excluded in total. Whole-brain average re-referencing was performed. Independent component analysis (ICA) with the extended Infomax algorithm was used to decompose EEG signals and artifacts. Ocular and muscular artifacts were identified and removed manually via ICLabel (Ezzatdoost et al., 2020). Finally, trials with amplitudes exceeding ±100 μV were excluded. After preprocessing, data from four participants were eliminated due to severe noise contamination and poor signal quality. The EEG recording data of the remaining 20 participants were retained for analyses.

2.6. EEG feature extraction

All EEG feature extraction and visualization were implemented using Python 3.9.

2.6.1. Time-domain analysis

The brain was divided into six brain regions: frontal lobe, central region, parietal lobe, occipital lobe, left temporal lobe, and right temporal lobe. The mapping between electrodes and regions is shown in Fig. A2 and Table A1. Five typical time-domain features, including mean, variance, standard deviation, root mean square (RMS), and peak-to-peak (PTP) amplitude, were extracted from single-channel EEG signals (Bastos-Filho et al., 2012). Features recorded from all electrodes within each brain region were averaged to obtain regional representative values. The extracted features of the six brain regions were concatenated to construct a 30-dimensional time-domain feature vector, followed by Z-score standardization. In addition, the MNE-Python toolbox was adopted to generate topographic maps under different conditions for visual analysis.

2.6.2. Frequency-domain analysis

Frequency-domain feature extraction was performed with NumPy and SciPy. Fast Fourier Transform (FFT) was applied to each epoch to calculate the power spectral density (PSD). The area under the curve (AUC) of five canonical frequency bands, namely δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-45 Hz), was computed as absolute power features (Schubring and Schupp, 2021). Based on the above indicators, five band power ratios, including θ/β, α/β, θ/α, δ/β, and β/γ, were calculated as relative power features (Jenke et al., 2014). Three advanced spectral features, namely spectral entropy, peak frequency, and spectral centroid, were further extracted (Tenev et al., 2025; Vidaurre et al., 2023). Alpha asymmetry features based on the left and right frontal lobes were also included (Aldayel et al., 2021), calculated according to the formula proposed by Touchette and Lee (2017), as shown below:

$$AWindex=\frac{\bar{\alpha }\left(Fp2,AF4,F8,F4\right)-\bar{\alpha }\left(Fp1,AF3,F7,F3\right)}{\bar{\alpha }\left(Fp2,AF4,F8,F4\right)+\bar{\alpha }\left(Fp1,AF3,F7,F3\right)}$$

All frequency-domain features were standardized within each participant and finally concatenated into a 14-dimensional frequency-domain feature vector.

2.7. Data analysis

Spearman's rank correlation analysis was performed to examine the correlations between each EEG feature and hedonic rating. The correlation coefficient r and p-value were calculated, and false discovery rate (FDR) correction was applied to account for multiple testing. Only feature pairs with FDR-adjusted p < 0.05 were screened as statistically significant. Independent-sample t-tests were primarily adopted for intergroup frequency-band comparisons between low-liking and high-liking conditions; considering the right-skewed distribution and outliers observed in normalized PSD values, Mann–Whitney U tests were further conducted to verify the robustness of group differences, and FDR correction was applied to both parametric and non-parametric test results. Meanwhile, to explore the overall association between the two sets of variables, canonical correlation analysis (CCA) was adopted to extract the first canonical variable of EEG features and hedonic ratings, and the canonical correlation coefficient as well as its significance level was further calculated.

2.8. Construction and evaluation of machine learning models

Hedonic rating classification models were constructed using Python 3.9 and scikit-learn. Labels were derived from overall liking scores to categorize samples into high-liking and low-liking groups for supervised binary classification. The median score of 4.8 was adopted as the cut-off threshold: trials with scores below 4.8 were defined as Cond 0 (low-liking), and trials with scores equal to or higher than 4.8 were defined as Cond 1 (high-liking). Notably, three repeated EEG recordings collected for each participant-sample unit were averaged prior to model construction, resulting in only one feature vector per participant–whisky combination. The dataset was split into training, validation, and test sets with a ratio of 6:2:2 while maintaining balanced class distributions, and no unit overlapped across the three subsets. Four classical classification models were established, including Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Random Forest (RF). A 5-fold cross-validation (GridSearchCV) was applied to automatically optimize the key hyperparameters. To ensure the reliability of the results, all experiments were independently repeated 100 times, and the average accuracy was used as the core evaluation metric. Simultaneously, the model performance was comprehensively evaluated by plotting receiver operating characteristic (ROC) curves and confusion matrices, as well as calculating precision, recall, and F1-score. Finally, the contributions of key features to the classification decisions were determined based on the feature weights from SVM and LR, and the feature importance from RF.

3. Results

3.1. The results of sensory scale

As shown in Fig. 2A, significant differences in consumer liking were observed among samples for all sensory attributes (p < 0.05). For pungency, Samples 4, 11, and 12 obtained relatively high scores, whereas Samples 1, 2, and 7 received lower ratings. In terms of aroma, Samples 7 and 8 achieved higher scores, while Samples 4 and 6 were rated lower. For taste, Samples 8 and 10 had the highest scores, with Sample 6 showing the lowest value. Regarding overall liking, Samples 8, 10, 11, and 12 received favorable evaluations, whereas Samples 1 and 6 had poor sensory acceptance. Overall, consumers showed higher preference for Japanese and Irish whisky, and lower liking for American and Scotch whisky. Inter-rater reliability of consumer hedonic ratings was assessed via ICC (Table A2). The ICC values for each liking attribute were as follows: pungency liking = 0.584, aroma liking = 0.711, taste liking = 0.805, overall liking = 0.536. This range of agreement is expected for untrained consumer subjects due to inherent individual differences in subjective flavour preference.

Fig. 2.

Hedonic scores and JAR ratings. (A) Hedonic scores for for pungency, aroma, taste, and overall liking, Different lowercase letters above bars indicate significant differences among samples (p < 0.05); (B) JAR ratings for pungency; (C) JAR ratings for aroma; (D) JAR ratings for taste. Stacked bars show the percentage of respondents rating each sample as “Too Little”, “JAR”, or “Too Much”.

As shown in Fig. 2B–D, Sample 10 presented the highest proportion of JAR responses for pungency (65%), Sample 7 for aroma (69%), and Samples 4 and 8 for taste (56%), indicating that these samples matched consumers’ ideal sensory intensity. In contrast, Sample 6 had the highest percentage of “Too Much” responses in aroma and taste, suggesting that most consumers perceived its flavor intensity as excessive. These results were consistent with the hedonic scoring trends.

Penalty analysis further illustrated the negative effects of deviations from the JAR level in each sensory attribute on overall liking (Fig. 3). The horizontal axis represents the proportion of consumers who gave non-JAR ratings, and the vertical axis indicates the mean drop in overall liking caused by such sensory deviations. The 20% dashed line denotes the critical industry threshold for sensory improvement (Zhi et al., 2016). For pungency, Samples 2 and 7 showed a “Too Much” proportion above 55%, accompanied by a mean drop exceeding 0.25, which should be regarded as high-priority sensory defects. In the aroma dimension, the “Too Much” ratio of Sample 6 exceeded 80%, with a corresponding mean drop of 0.61, and Sample 10 also reached over 70%. For taste, Sample 6 had a “Too Much” proportion higher than 60% and a mean drop of 0.15, while Samples 3 and 9 also exceeded 40%. In general, Sample 6 exhibited high deviation ratios and considerable mean drops in aroma and taste, reflecting severe imbalance in flavor intensity. By comparison, Samples 8 and 10 maintained low deviation proportions and minor mean drops across all attributes, demonstrating stable and well-controlled sensory intensity. These findings are consistent with the JAR evaluation and hedonic score results.

Fig. 3.

Penalty analysis plot. (A) Penalty analysis plot for pungency; (B) Penalty analysis plot for aroma; (C) Penalty analysis plot for taste. The x-axis shows the percentage of consumers who rated each sensory attribute as deviating from the JAR level, and the y-axis represents the corresponding mean drop in overall liking.

3.2. Machine learning models based on raw EEG data

Machine learning-based classification and prediction were directly performed on the preprocessed EEG data. The downsampled data containing 32 channels and 150 time points were flattened into a 4800-dimensional raw feature vector. Four classical classification models were constructed using LR, RF, KNN, and SVM, The corresponding classification performance is presented in Table 2. The results indicated that the classification accuracy of the LR, RF, KNN and SVM models reached 51.0% ± 7.1%, 51.4% ± 7.3%, 50.4% ± 7.0% and 50.6% ± 6.8%, respectively. All models yielded performance close to the random guessing level, indicating limited effective predictive ability (Çorbacıoğlu and Aksel, 2023). This finding is consistent with previous studies (Fu et al., 2025; Zhao et al., 2026).

Table 2.

Classification performance of four models using raw EEG data.

	Logistic Regression			Random Forest			KNN				SVM
	precision	recall	f1-score	precision	recall	f1-score	precision	recall	f1-score	precision	recall	f1-score
Cond 0	0.52	0.51	0.52	0.52	0.53	0.53	0.52	0.51	0.51	0.52	0.51	0.51
Cond 1	0.50	0.51	0.51	0.50	0.49	0.50	0.49	0.50	0.50	0.49	0.51	0.50
accuracy			0.51			0.51			0.50			0.51
macro avg	0.51	0.51	0.51	0.51	0.51	0.51	0.50	0.50	0.50	0.51	0.51	0.51
weighted avg	0.51	0.51	0.51	0.51	0.51	0.51	0.50	0.50	0.50	0.51	0.51	0.51

Precision: The proportion of samples correctly belonging to a specific category among all samples predicted as that category by the model.

Recall: The proportion of samples correctly identified by the model among all actual samples of a given category.

F1-score: The harmonic mean of precision and recall, used to comprehensively evaluate the classification performance of a single category.

Cond 0: dislike condition.

Cond 1: like condition.

Accuracy: The overall proportion of correctly classified samples across all categories.

Macro avg: The arithmetic mean of each metric calculated independently for all categories.

Weighted avg: The average value of each metric weighted by the sample size of each category.

3.3. Time-domain and frequency-domain features of EEG

3.3.1. Time-domain features

As illustrated by EEG topographic maps (Fig. 4), EEG activity under both conditions showed obvious temporal evolution trends and spatial distribution discrepancies. Significant differences were observed during the middle and late stages (2-4 s): the activation intensity of the central region was higher in Cond 1 than in Cond 0. Meanwhile, Cond 1 exhibited a more concentrated and stronger activation pattern, with relatively stable activation sustained in the late stage (4-5 s). In contrast, neural responses of Cond 0 were more transient and spatially dispersed. These findings demonstrate that differences in beverage preference can be clearly reflected in the temporal dynamics and spatial features of EEG signals.

Fig. 4.

Second-by-second averaged EEG topographic maps within the 0–5 s time window under Cond 0 and Cond 1. Cond 0: dislike condition; Cond 1: like condition.

3.3.2. Frequency-domain features

This study performed within-subject standardization on the AUC of PSD across five frequency bands: δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-45 Hz). Independent-sample t-tests were used for statistical comparison; complementary Mann–Whitney U tests were additionally conducted to verify result robustness given the right-skewed distributions and outliers observed in normalized PSD data, and Benjamini–Hochberg FDR correction was applied to all resulting p-values. The results are shown in Fig. 5A–E. Except for the γ band, significant differences were detected between preference groups in all other bands (p < 0.05) (Table A3). The α band exhibited the most prominent intergroup difference (p < 0.001). In Cond 1, the normalized indicators were generally positively distributed with a median above zero; in Cond 0, the distribution tended to be negative with a median below zero. This result indicates that increased α band power is closely related to higher sensory liking. The β band also showed significant group differences (p = 0.004), with overall distribution values higher in Cond 1 than in Cond 0. Notably, α, β and θ showed consistent significant results in both tests, confirming these band differences are robust against non-normal data and outliers. In contrast, δ was only significant in parametric analysis. These findings demonstrate that power changes in the α and β bands act as potential predictive indicators for distinguishing liking and disliking states. Correlations among all frequency bands are presented in Fig. 5F. All bands showed positive correlations. The highest correlation occurred between the β and γ bands (r = 0.91), followed by δ–θ (r = 0.77) and α–β (r = 0.65). The δ and γ bands showed the lowest correlation (r = 0.45).

Fig. 5.

Comparisons of normalized AUC of PSD across frequency bands and their correlations between liking conditions. (A–E) Boxplots showing normalized AUC values for δ, θ, α, β, and γ bands in Cond 0 (dislike) and Cond 1 (like) (*, p < 0.05, **, p < 0.01, ***, p < 0.001; ns, not significant). (F) Correlation matrix of the five frequency bands, the color bar represents the correlation coefficient r values.

3.4. Correlation analysis between EEG and sensory liking

To explore the intrinsic relationships between EEG features, sensory liking and intensity ratings, 14 frequency-domain features and 30 time-domain features were extracted. These features were selected based on previous EEG studies demonstrating their relevance to emotional and affective processing (detailed in Section 2.6). The full list of features with corresponding identifiers is provided in Table A4. Spearman correlation matrices between these features and hedonic indicators were calculated, and Benjamini–Hochberg FDR correction was applied to all resulting p-values. CCA was adopted to examine the overall correlation patterns. The CCA results revealed a significant canonical correlation between EEG features and liking dimensions (r = 0.477, p < 0.001), proving that EEG signals can effectively reflect variations in sensory evaluation. Fig. 6 visualizes the correlations between 44 EEG features and 7 liking dimensions, with distinct differences observed in the contribution of individual features. After FDR correction, only associations linked to Liking 1 (pungency liking) and Liking 4 (overall liking) remained statistically significant, indicating these two hedonic dimensions were most closely modulated by EEG signals. Feature 14 showed the strongest positive correlation with Liking 4 (r = 0.359, p < 0.001), acting as a primary predictive indicator for discriminating this dimension; features 3, 8 and 40 also had significant correlations with Liking 4. Features 26–28, including parietal lobe variance, standard deviation and RMS, were significantly negatively correlated with Liking 1. This trend indicates that greater activity fluctuations, higher energy and larger peak differences in the parietal lobe correspond to lower subjective sensory liking.

Fig. 6.

Correlation plot. Detailed information on each numbered EEG feature is provided in Table A4; Liking 1: Pungency liking; Liking 2: Aroma liking; Liking 3: Taste liking; Liking 4: Overall liking; Liking 5: Pungency intensity; Liking 6: Aroma intensity; Liking 7: Taste intensity. *p < 0.05, **p < 0.01, ***p < 0.001.

3.5. Machine learning models based on EEG features

Since previous studies have demonstrated that frequency-domain features outperform time-domain features in hedonic recognition (Jenke et al., 2014), this study selected all frequency-domain features together with 6 representative time-domain features as model inputs. These 6 time-domain features correspond to the mean voltage values of six brain regions, which can effectively reflect overall neural activity. Four classification models were constructed and compared, including LR, RF, KNN and SVM. The results showed that the RF model achieved the best comprehensive performance, with a training accuracy of 98.2%, a validation accuracy of 82.1%, and a test accuracy of 82.6% ± 4.8% (mean ± SD). As illustrated by the confusion matrix (Fig. 7B) and classification report (Table 3), the recall for the dislike condition (Cond 0) was 81%, and the recall for the liking condition (Cond 1) reached 84%. The F1-score of the two categories was 0.82 and 0.83, respectively. These outcomes indicated that the model maintained balanced recognition ability for positive and negative samples, with favorable discrimination and stability. The performance of the remaining models was ranked as follows. The SVM model yielded a training accuracy of 87.1%, a validation accuracy of 80.2%, and a test accuracy of 81.1% ± 5.5%. The LR model obtained a training accuracy of 86.9%, a validation accuracy of 79.5%, and a test accuracy of 80.8% ± 5.5%. The KNN model presented the weakest performance, with a training accuracy of 84.4%, a validation accuracy of 70.8%, and a low test accuracy of 70.5% ± 6.5%.

Fig. 7.

Comparison of classification results across different models. (A–D) Confusion matrices of LR, RF, KNN and SVM. The vertical axis denotes the true label, and the horizontal axis represents the predicted label. The top-left, top-right, bottom-left and bottom-right sections correspond to TN, FP, FN and TP, respectively. (E) ROC curves and the corresponding AUC of the four classification models.

Table 3.

Classification performance of four models using EEG features.

	Logistic Regression			Random Forest			KNN				SVM
	precision	recall	f1-score	precision	recall	f1-score	precision	recall	f1-score	precision	recall	f1-score
Cond 0	0.80	0.81	0.81	0.83	0.81	0.82	0.71	0.67	0.69	0.81	0.81	0.81
Cond 1	0.82	0.80	0.81	0.82	0.84	0.83	0.70	0.74	0.72	0.81	0.81	0.81
accuracy			0.81			0.83			0.71			0.81
macro avg	0.81	0.81	0.81	0.83	0.83	0.83	0.71	0.70	0.70	0.81	0.81	0.81
weighted avg	0.81	0.81	0.81	0.83	0.83	0.83	0.71	0.71	0.70	0.81	0.81	0.81

Feature weight analysis was conducted on the models (Fig. 8), and the results indicated that different features contributed unequally to classification. Feature 14 (frontal lobe α asymmetry) and Feature 3 (AUC of the α band) presented high positive weights, which served as primary predictive indicators for identifying the liking state. As a core component of the approach-avoidance motivational index, elevated frontal lobe α asymmetry indicates greater left frontal activation, and it is closely associated with the processing of positive approach motivation and pleasant emotions (Aldayel et al., 2021). α waves are linked to wakeful relaxation and eyes-closed resting states (Wang et al., 2011). Increased α power is directly correlated with subjective perceptions of pleasure, comfort and low stress, and a high level of α power reflects that participants remained in a relaxed and pleasant state during tasting. In comparison, Feature 7 (α/β power ratio) and Feature 2 (AUC of the θ band) showed distinct negative weights. The α/β power ratio, as a fundamental frequency-domain feature, has been used to differentiate arousal levels associated with different emotions (Jenke et al., 2014). In contrast, findings regarding θ band power as a predictor have been inconsistent across studies, and the most reliable EEG markers for preference prediction are frontal alpha asymmetry (FAA) and late positive potential (LPP) (Byrne et al., 2022). Notably, three models including LR, SVM and RF shared highly consistent key features. This finding confirms that these features are not random results of model fitting, but candidate predictive EEG signals closely associated with subjective preference. Overall, the differences in feature importance not only support the interpretability of the classification model, but also verify that EEG signals contain core features steadily associated with subjective preference, which can provide potential electrophysiological biomarkers for further research.

Fig. 8.

Feature weight and importance ranking across different models. (A) Feature weights of the LR model; (B) Feature weights of the SVM model; (C) Feature importance of the RF model. Green and orange bars represent positive weight or importance, while red bars indicate negative weight.

4. Discussion

4.1. Sensory evaluation

Sensory evaluation revealed clear preference divergence across samples: Japanese and Irish whisky gained higher overall liking, whereas American and Scotch samples were less favored. Penalty analysis identified Sample 6 with excessive flavor intensity as the least accepted product.

4.2. EEG feature analysis

Topographic maps showed stronger and more sustained activation over the central region under Cond 1. Anatomically, the postcentral gyrus (PCG) serves as the deep functional source of the central region; it integrates gustatory inputs and relays signals to the insula to mediate beverage preference evaluation (Yang et al., 2023). Significant intergroup differences existed in δ, θ, α and β bands, with α and β bands presenting the most prominent intergroup separation. The α band typically reflects brain resting state, inhibitory control and cortical arousal, and its enhancement is associated with comfort, pleasure, focused attention and relaxation (Wang et al., 2011). The β band is closely linked to wakefulness, vigilance, active cognition, sustained attention and emotional arousal, and serves as a core indicator for external stimulus processing and decision judgment (Ray and Cole, 1985). In addition, although significant differences were also observed in the δ and θ bands, these two bands are generally considered to have relatively weak discrimination ability for emotion and preference evaluation, and are less important than other frequency bands (Wang et al., 2011). Spearman correlation results confirmed significant correlations between EEG features and sensory hedonic scores. Frontal alpha asymmetry showed a positive correlation with overall liking, and parietal time-domain features exhibited a negative correlation with pungency liking.

4.3. Machine learning classification performance

Models trained on raw EEG data only reached an accuracy of around 51%, equivalent to random guessing. A plausible explanation is that raw EEG signals still contain substantial noise components irrelevant to beverage preference, even after preprocessing (Chaddad et al., 2023). Such noise severely reduces feature learning efficiency and classification accuracy. Accordingly, the direct utilization of raw EEG data cannot achieve effective discrimination of individual preference. It is necessary to extract refined EEG features to eliminate noise interference and capture valid neural representations, thereby improving the overall performance of classification models.

Compared with models based on raw EEG signals, the models built on extracted features in this study exhibited substantially higher performance, confirming that EEG features are more suitable as machine learning inputs for preference prediction. In addition, the classification accuracy achieved in this study was higher than those reported in previous alcoholic beverage preference studies using EEG and machine learning, including 68.69% accuracy for Baijiu-food pairing preference prediction (Fu et al., 2025) and 64%–72% accuracy for beer liking prediction (Gonzalez Viejo et al., 2019). Notably, these prior studies also measured other biometric signals beyond EEG, such as facial expressions and heart rate, which provides valuable implications for future work.

Nevertheless, the prediction performance of the classification models in this study remains relatively limited, which may be attributed to two main factors. On the one hand, a wide variety of whisky samples were adopted in the experiment. Differences in flavor, physicochemical properties and other dimensions across samples triggered complex EEG responses, which may mask weak neural signals associated with subjective preference to a certain extent. This makes it difficult for the models to efficiently extract stable and representative preference-related features. On the other hand, model labels were established based on hedonic scores. Such subjective ratings are inherently unstable and susceptible to confounding factors including emotion, environment and transient individual states (Mukhopadhyay and Chung, 2016). Label noise further reduced the machine learning algorithm efficiency and generalization ability for preference.

4.4. Limitations

In addition, several limitations of the study should be acknowledged. First, the sustained delivery of taste stimuli hindered the extraction of typical ERP waveforms, resulting in insufficient interpretation of time-domain features. Second, the sample size is relatively small, and participant representativeness is limited; all participants were young college students within a narrow age bracket, restricting the external validity of our results. Third, individual differences existed in neural responses across participants. For instance, individuals with stronger baseline activity in the left parietal and occipital regions tend to report higher hedonic ratings for food (Van Bochove et al., 2016). Even after intra-individual standardization, constraints such as sample size still limited the generalizability of the research findings. Future studies may focus on improving the analysis of long-duration EEG waveforms that are compatible with natural and prolonged tasting processes, in an attempt to obtain typical ERP components under real drinking conditions. Expanding the sample size and recruiting participants with more diverse demographic characteristics will help reduce the impact of individual differences and further improve the generalizability and robustness of the findings. Furthermore, integrating EEG with other multimodal physiological signals, such as heart rate and facial expressions, is recommended to serve as input features for machine learning, so as to establish a more comprehensive and reliable system for the objective evaluation of consumer preference.

5. Conclusion

This study combined subjective sensory evaluation and objective physiological signals to systematically analyze consumer preference differences for whisky and the corresponding neural activity patterns. Machine learning models were constructed based on EEG features to distinguish hedonic states. The results revealed significant discrepancies in consumer ratings of pungency, aroma, taste and overall liking across samples, with Japanese and Irish whisky receiving higher preference. EEG findings demonstrated that high preference was accompanied by stronger and longer-lasting activation in the central region, as well as marked increases in α and β band power. Features including frontal lobe α asymmetry and parietal lobe activity stability are potential predictive indicators of consumer preference. Machine learning outcomes indicated that raw EEG signals failed to effectively discriminate hedonic states, while feature extraction greatly improved model performance. The RF model achieved the best performance, and core features exhibited high consistency across multiple algorithms. Future research should adopt broader sample coverage and multimodal physiological indicators to further enhance model generalizability. This approach may provide novel insights and technical support for the objective evaluation and flavor improvement of beverages.

Ethics declaration

Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. This study was approved by the Ethics Committee of Jiangnan University. (Approval No. JNU202506RB010)

Data availability

Data will be made available on request.

Declaration of competing interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Handling Editor: Professor Georgios Leontidis

Appendix A. Supplementary data

The following is the supplementary data to this article.