DEAP Dataset: Emotion Analysis & Dark Matter Studies

• DEAP Dataset comprises two distinct resources: one for affective computing with EEG and peripheral recordings for emotion labeling, and one for dark matter studies with PMT waveform data.
• In affective computing, DEAP involves controlled emotion elicitation, robust preprocessing, and protocol benchmarking to assess valence-arousal dimensions.
• In rare-event physics, DEAP-1 and DEAP-3600 underpin background modeling, isotope measurements, and sensitivity studies for detecting dark matter events.

The term DEAP Dataset is used for two technically unrelated resources in contemporary research. In affective computing, DEAP denotes a benchmark dataset for emotion analysis from EEG and peripheral physiological signals, organized around stimulus-evoked valence–arousal annotation. In rare-event physics, DEAP refers to event datasets produced by the DEAP liquid-argon dark-matter program at SNOLAB, especially DEAP-1 and DEAP-3600, where the fundamental data are PMT waveforms used for pulse-shape discrimination, background decomposition, isotope measurements, and sensitivity studies (Khan et al., 4 Aug 2025, Amaudruz et al., 2012).

1. Dual usage and domain disambiguation

The two usages differ at the level of data structure, inference target, and validation protocol.

Usage of “DEAP”	Core data unit	Primary research use
Affective-computing DEAP	Trial-level EEG and peripheral recordings with affect labels	Valence/arousal classification, regression, protocol benchmarking
DEAP-1 / DEAP-3600	Event-level PMT waveforms in liquid argon	Background modeling, PSD studies, isotope measurements, rare-event sensitivity

In the affective-computing literature, DEAP is a public benchmark centered on physiological recordings acquired during emotion elicitation. In the dark-matter literature, DEAP is an experimental program using single-phase liquid argon, and the corresponding “dataset” is the detector output itself. This distinction is essential because identical terminology masks incompatible notions of sample independence, label formation, calibration, and uncertainty propagation (Khan et al., 4 Aug 2025, Collaboration et al., 22 Jan 2025).

2. Affective-computing DEAP: acquisition, preprocessing, and labels

DEAP is described as a benchmark dataset for emotion analysis with EEG and peripheral physiological signals. It contains recordings from 32 healthy participants watching 40 one-minute music videos intended to elicit affective responses. The raw data are reported as 48 channels per participant, including 32 EEG, 12 peripheral, three unused, and one status channel, recorded at 512 Hz. The preprocessed data are downsampled to 128 Hz, with EOG artifacts removed, 4–45 Hz band-pass filtering, common average reference, and 60-second trials plus 3-second baseline. After each video, participants rated valence, arousal, dominance, and liking on a 9-point scale; later work places these labels within Russell’s circumplex model, with valence and arousal as the primary axes (Khan et al., 4 Aug 2025).

A later methodological replication emphasizes the dataset’s single-trial structure and reports 40 stimuli per participant, balanced across the four valence–arousal quadrants HAHV, LAHV, LALV, and HALV, with 10 videos each. The same study notes that the original DEAP analysis used Naïve Bayes, LOVO cross-validation, and reported accuracy and macro-averaged F1 with one-sample t-tests for statistical significance (Rahmani et al., 2024).

Reduced-channel EEG work operationalizes the dataset as 40 measurement channels per trial: 32 EEG channels plus 8 additional physiological signals, including hEOG, vEOG, EMG, GSR, BVP, respiration, and temperature. For SAM-based experiments, high versus low valence or arousal is obtained by thresholding at 5 on the 1–9 scale. This later representation is consistent with the raw-versus-measurement distinction, because the larger raw channel count explicitly includes unused and status channels (Hoffsommer et al., 1 Oct 2025).

3. Evaluation protocols, leakage modes, and methodological controversy

A 2025 review of 101 DEAP-based studies argues that the dominant methodological problem is not model family selection but evaluation validity. The review identifies widespread failures involving segmentation leakage, global feature selection, test-set-driven hyperparameter tuning, neglect of class imbalance, and insufficient reporting, and states that approximately 87% of reviewed papers contained one or more such issues. Only 38 of 101 papers reported imbalance-aware metrics, and only 8 provided code links (Khan et al., 4 Aug 2025).

The most consequential issue is improper segmentation or windowing. When short windows from the same 60 s trial are distributed across train and test folds, temporal autocorrelation, baseline structure, noise characteristics, and stimulus-specific patterns leak across partitions. In the review’s DEAP valence experiment, 5-fold CV over all segments rose from approximately 53% accuracy without segmentation to above 90% with 6 segments per trial and nearly 100% with 60 segments per trial, whereas a correct leave-one-trial-out protocol remained near 52% regardless of segmentation level. The same paper reports that feature-selection leakage produced inflation of roughly 14–18 percentage points, and incorrect hyperparameter tuning produced inflation of roughly 21–27 percentage points. These effects are large enough to reorder published leaderboards (Khan et al., 4 Aug 2025).

The paper therefore recommends grouped CV, LOTO for subject-dependent settings, LOSO for subject-independent settings, and nested CV for feature selection and hyperparameter search. For imbalanced binary tasks it emphasizes balanced accuracy,

$BA = \frac{1}{2}(TPR + TNR)$

in contrast to plain accuracy,

$Acc = \frac{TP + TN}{TP + TN + FP + FN}.$

It also reports selective comparison bias: in 21 binary-classification papers with comparison tables, reported performance correlated strongly with the highest prior result chosen for inclusion, with Pearson $r = 0.78$ for valence and $r = 0.79$ for arousal, while only 4 papers acknowledged stronger prior results than their own (Khan et al., 4 Aug 2025).

4. Replication, transfer, and reduced-sensor modeling

A direct replication of the DEAP single-trial peripheral-signal pipeline uses the preprocessed Python pickle package, includes all subjects and trials, and preserves the original signal family: BVP, RSP, EDA, SKT, EMG, and EOG. The preprocessing mirrors the earlier DEAP study: BVP and EDA are detrended by subtracting a 256-point moving average; RSP is band-pass filtered at 0.15–0.35 Hz; EMG and EOG are band-pass filtered at 4–40 Hz; SKT is not preprocessed. Feature selection is performed with Fisher’s linear discriminant score,

$$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$

using an empirical threshold of 0.3, followed by Gaussian Naïve Bayes, subject-dependent LOVO CV, and averaging of accuracy and macro-F1 across subjects. The replication reports valence accuracy $0.631$, valence F1 $0.606$ $(p<0.01)$, arousal accuracy $0.616$, and arousal F1 $0.542$ $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$0, closely matching the original DEAP peripheral-signal results $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$1 for valence and $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$2 for arousal (Rahmani et al., 2024).

The same work then treats DEAP as a methodological template for transfer to wearable sensing. Its role is not merely comparative; it provides the exact pipeline whose replication is used to validate later SCG and ADR analyses. This makes DEAP a protocol anchor as much as a dataset (Rahmani et al., 2024).

A separate ViT study uses preprocessed DEAP end-to-end for reduced-channel EEG classification. Each channel is normalized, transformed with a Continuous Wavelet Transform into a scaleogram, resized to $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$3, and processed by a Linformer-based Vision Transformer. All 32 participants and all 40 trials per participant are included. The main 5-fold CV split is explicitly not dependent on participant ID, so samples from the same person can appear in both train and test folds. Under this protocol, the Emotiv subset achieves mean four-class accuracy $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$4 with best fold $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$5, EEG-only reaches mean $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$6, Muse-12 reaches $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$7, and hEOG alone reaches mean $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$8; for hEOG, an additional cross-person experiment reports mean accuracy around $Acc = \frac{TP + TN}{TP + TN + FP + FN}.$9 when six subjects are held out per fold. The same study also shows that label choice matters: under SAM rather than VAQ_Estimate, the Emotiv subset drops to mean $r = 0.78$0, and 671 out of 1280 scaleograms change quadrants (Hoffsommer et al., 1 Oct 2025).

These results show that DEAP supports both faithful pipeline replication and aggressive channel reduction. They also suggest that reported performance must always be read together with the split design, the label construction rule, and the distinction between subject-mixed and strict cross-person evaluation.

5. DEAP-1: prototype detector dataset and background taxonomy

In rare-event physics, DEAP-1 was a $r = 0.78$1 kg single-phase liquid-argon scintillation detector operated underground at SNOLAB to develop and validate background-rejection techniques, materials, and analysis methods for DEAP-3600. Its active argon volume was a 28 cm long, 15 cm diameter acrylic cylinder with a $r = 0.78$2 kg active target; the inner surfaces were coated with a 1–4 $r = 0.78$3m TPB film, and two PMTs viewed the volume through light guides. The detector digitized PMT waveforms over $r = 0.78$4s. The principal PSD observable was

$r = 0.78$5

with electron recoils clustering around $r = 0.78$6 and nuclear recoils around $r = 0.78$7–$r = 0.78$8. The electron-equivalent scale was set with $r = 0.78$9Am $r = 0.79$0 keV and $r = 0.79$1Na $r = 0.79$2 keV gammas, yielding $r = 0.79$3, $r = 0.79$4, and $r = 0.79$5 PE/keV at $r = 0.79$6 keV for G1, G2, and G3, respectively. With a quenching factor $r = 0.79$7, a 30–50 keVee ROI corresponds to approximately 120–200 keV nuclear recoil energy (Amaudruz et al., 2012).

The central purpose of DEAP-1 was background decomposition. Full-energy alpha decays from radon chains provided direct access to $r = 0.79$8Rn and $r = 0.79$9Rn rates, and the measured $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$0Rn decay rate in liquid argon was stable at 16–26 $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$1Bq/kg. Low-energy nuclear-recoil-like events were attributed to three sources: radon daughters on active surfaces, electromagnetic events misidentified because of PSD inefficiency, and leakage from outside the fiducial volume due to imperfect position reconstruction. Two mechanisms were especially important for alpha-related leakage. “Geometric alphas” arose in low light-collection regions near the windows and neck, while “surface alphas” on TPB or acrylic could produce reduced apparent energy according to

$$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$2

For users of a DEAP-like detector dataset, the core high-level quantities are digitized waveforms, TotalPE, $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$3, and axial position $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$4; below approximately 50 keVee, the G3 fiducial spectrum was described by PSD leakage plus window leakage, whereas 50–200 keVee was dominated by surface alphas from radon daughters on TPB (Amaudruz et al., 2012).

6. DEAP-3600: electromagnetic background, isotope, and extended-signal datasets

DEAP-3600 is a 3.3 tonne single-phase liquid-argon detector located 2.1 km underground at SNOLAB, with a spherical acrylic vessel of radius 85 cm, a 3 $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$5m TPB coating, and 255 Hamamatsu R5912 PMTs. An open 247.2 d dataset, analyzed without fiducial cuts, was used to construct a detailed electromagnetic-background model over $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$6 from 2000 to 35000, corresponding to approximately 290 keV to 5 MeV. The PSD variable remained $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$7 with a 150 ns prompt window and a 10 $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$8s total window, with typical averages $$J(f) = \frac{|\mu_1 - \mu_2|}{\sigma_1^2 + \sigma_2^2},$$9 for NR and $0.631$0 for ER. The background model included internal $0.631$1Ar, $0.631$2Ar/$0.631$3K, radon daughters, surface $0.631$4Pb/$0.631$5Bi, and external components from the acrylic vessel, light guides, filler blocks, PMTs, stainless steel shell, and neutron-capture gammas. A Bayesian Analysis Toolkit MCMC fit yielded a specific atmospheric $0.631$6Ar activity of $0.631$7Bq/kg of argon (Ajaj et al., 2019).

A different DEAP-3600 dataset, spanning 2016–2020, was used for a direct $0.631$8Ar half-life measurement. The detector contained $0.631$9 kg of atmospheric liquid argon; PMT waveforms were digitized for $0.606$0s, and a prescale of 100 was applied in the low-$0.606$1 electron-recoil region, leaving roughly $0.606$2 $0.606$3Ar decays per 24 hours after prescaling. The half-life analysis selected 700–1200 PE, approximately 115–195 keVee, with $0.606$4, $0.606$5, and a $0.606$6s deadtime cut to suppress late-light contamination. The rate model explicitly included single, double, and triple $0.606$7Ar pile-up, $0.606$8Ar–Cherenkov pile-up, and a constant ER background, fitted in weekly bins. The result was

$0.606$9

with starting $(p<0.01)$0Ar rate $(p<0.01)$1 Hz (Collaboration et al., 22 Jan 2025).

DEAP-3600 data have also been repurposed for non-WIMP feasibility studies. A 2026 analysis of supermassive charged gravitinos models the signal as an ER-like, straight-line excitation pattern through liquid argon, producing broadly distributed PMT light over approximately $(p<0.01)$2s, $(p<0.01)$3 in the range 0–0.1, and total observed $(p<0.01)$4 of order $(p<0.01)$5–$(p<0.01)$6. Under the study’s assumptions, the expected rate in DEAP-3600 is 0.15 events in 820 days, so the present exposure is interpreted as allowing only exclusion in the high-velocity regime rather than discovery (Olszewski, 23 Apr 2026).

7. Scientific significance and recurrent misconceptions

In affective computing, DEAP remains a central benchmark, but its role is increasingly methodological rather than merely archival. Exceptionally high accuracy claims cannot be interpreted in isolation from grouped dependencies, thresholding rules, feature-selection scope, and whether cross-validation is performed over segments, trials, or subjects. The later review literature therefore treats protocol specification as part of the dataset definition itself (Khan et al., 4 Aug 2025).

In rare-event physics, DEAP datasets are not fixed benchmark tensors but calibrated detector records whose meaning depends on light-yield monitoring, PSD parameterization, event reconstruction, fiducialization, and control samples. DEAP-1 established that surface alphas, EM PSD leakage, and fiducial leakage statistically account for the observed ROI events; DEAP-3600 extended this framework to full electromagnetic background decomposition and isotope metrology, including direct determinations of $(p<0.01)$7Ar activity and $(p<0.01)$8Ar half-life (Amaudruz et al., 2012, Collaboration et al., 22 Jan 2025).

A persistent misconception is that “DEAP dataset” denotes a single canonical resource. The literature instead supports a sharper distinction: in one field DEAP is a benchmark for affective state estimation, while in another it is a family of SNOLAB detector datasets for liquid-argon rare-event physics. Precision in that distinction is necessary because the two usages differ not only in modality and scale, but also in what counts as a valid split, a meaningful label, and a reproducible result.