Can 2D CNNs classify hypnotic susceptibility from minimally preprocessed EEG?

A binary decision between relaxation and mental effort

The study asks whether a convolutional neural network can tell, from EEG alone, whether a person is in a guided-imagery relaxation state or performing a demanding cognitive task. Sixty-two male participants aged 19–24 were recorded with a 256-channel dense-array cap at 250 Hz for 20 minutes each. Thirty listened to a guided-imagery relaxation recording; 32 performed mental retrieval tasks such as naming voivodeships or US states, were told they would later write down what they recalled, and worked under time limits. Both groups kept their eyes closed in a darkened room.

The label is therefore a group condition, not a clinical diagnosis of hypnotisability. The paper frames the target as hypnotic susceptibility, but the classifier distinguishes HIPNO from CONTROL recordings made during these two tasks. That distinction matters for reading every accuracy figure below: a high score means the network recovered task-group membership from the signal, not that it measured how hypnotisable an individual is.

Minimal preprocessing leaves the network to find structure

Preprocessing was deliberately light. Bad channels were detected with the RANSAC algorithm in the pyprep toolkit, removed and interpolated; each trial was bandpass filtered at 1–40 Hz and z-score normalised per channel using that channel's mean and standard deviation over the recording. No artifact removal and no manual feature extraction were applied. The authors' point is that filtering plus normalisation can be enough when a network learns from the raw signal directly.

The network treats each 10-second window as a two-dimensional matrix, with electrodes along one axis and time samples along the other. This layout lets 3×3 kernels respond to correlated activity across neighbouring electrodes within a short time span, which is why the design is called a 2D CNN rather than a 1D one. A second block doubles filter depth and pools spatial information, and a dropout rate of 0.50 in the dense layers compensates for the small number of training examples per window-level model.

The sliding-window framework is what produces the study's large comparison space: 124 ten-second windows across seven anatomically defined regions, giving up to 868 independently trained models. Each model is scored on one region and one window, so the paper can ask which cortical areas carry a discriminative signal and when.

Random-split screening favours posterior regions

In the first screening stage, each window–region combination was trained and tested with a segment-level 90%/10% random split and a fixed seed of 42. A window counted as discriminative if accuracy exceeded 0.67. The occipital region dominated: 112 of 124 windows (90.3%) passed, with peak accuracy 0.8026 and mean 0.7206. The parietal region followed with 95 of 124 windows (76.6%), peak 0.7730 and mean 0.7087.

Anterior regions produced far fewer passing windows but occasionally higher peaks. The prefrontal region held the single best window in the whole experiment at 0.8224, yet only 24 of 124 windows (19.4%) cleared the threshold, with mean 0.7086. Central and frontal regions passed 20 of 124 (16.1%) and 13 of 124 (10.5%) windows, peaking at 0.8026 and 0.7829. Temporal and facial regions produced no window above threshold.

This split is a screening device, not a generalisation estimate. Consecutive 50-sample sub-segments from the same 10-second window of the same subject can land in both training and test sets, so a classifier can learn a person's individual EEG fingerprint instead of a group difference. The authors state this limitation explicitly and restrict the random split to ranking windows.

Subject-independent testing reverses the ranking

Leave-one-subject-out cross-validation removes that leakage. In each of 62 folds, the model is trained from scratch on 61 subjects and evaluated on all segments from the held-out subject, so every score reflects performance on a person the network never saw. The best window per discriminative region was re-run this way.

The ranking changed sharply. The occipital region, dominant in screening, fell to mean accuracy 0.453 with AUC 0.429 — near chance. Parietal reached 0.631 (AUC 0.680), prefrontal 0.624 (AUC 0.681), central 0.591 (AUC 0.599), and frontal 0.646 (AUC 0.692). The authors read the occipital collapse as evidence of subject-specific EEG fingerprinting rather than a genuine group difference: the network had learned to recognise individuals, and that skill did not transfer.

Per-fold standard deviations were large, from 0.228 to 0.353, and the paper attributes this to a bimodal distribution in which individual held-out subjects were either classified well above chance or near chance. A mean accuracy in the 0.6s therefore describes an uneven population of subjects, not a uniform moderate skill.

Ensembling windows helps the prefrontal region most

The ensemble variant averages softmax probabilities across K independently trained models, each using a different top-ranked 10-second window of the same region, to produce one prediction per subject. With K = 10, the prefrontal region improved from a single-window mean of 0.624 to 0.738, a gain of 0.114, with AUC 0.761. Frontal rose by 0.021 to 0.667 (AUC 0.723) and central by 0.038 to 0.629.

Not every region benefited. Parietal dropped by 0.022 to 0.609, and occipital stayed near chance at 0.462 with AUC 0.472 regardless of ensemble size. The authors conclude that averaging windows cannot manufacture a generalisable signal where none exists. Comparing K = 3 with K = 10 also showed diminishing returns: prefrontal gained only 0.007 from the larger ensemble, while frontal, central and parietal did better with K = 3, which the authors attribute to less discriminative windows adding noise to the soft-vote average.

The practical reading is that window aggregation is a targeted fix for a region whose discriminability is real but intermittent. It worked where single windows were already above chance and failed where they were not.

Anterior markers generalise but fluctuate in time

The paper's central interpretation is a posterior–anterior split between consistency and transferability. Occipital and parietal regions gave the most temporally stable separation during screening, which the authors link to the role of visual imagery in guided imagery and to attentional and multisensory processes associated with parietal cortex. But that stability did not survive subject-independent testing.

Anterior regions showed the opposite profile: fewer discriminative windows, yet better cross-subject generalisation, with the frontal region highest under LOSO. The authors suggest frontal discriminability is transient, possibly reflecting episodic engagement of top-down executive control at particular moments rather than a sustained state marker. This is their explanation, not a tested mechanism; the study measures classification performance, not the cognitive process behind it.

The absence of any discriminative window in temporal and facial regions is also informative. Both groups listened to audio throughout, differing only in content, so the authors argue the group difference is not driven by differential auditory processing. The facial region, which mainly reflects peripheral muscle activity, was expected to contribute little.

What the accuracy numbers do not settle

Several boundaries limit how far these results travel. The sample is 62 male participants aged 19–24, recruited partly because the institute had few female students and long hair complicated cap placement; the authors acknowledge that the conclusions apply only to men. The comparison is between two task conditions, so the classifier's target is guided imagery versus cognitive effort, and the paper does not report a separate measure of individual hypnotisability to validate the label against.

The best subject-independent numbers are also modest in absolute terms. Prefrontal ensemble accuracy of 0.738 and frontal single-window accuracy of 0.646 sit above chance but well below the 0.82 seen in the leaky random split, and the high per-subject variance means some individuals were classified near chance. The paper compares its results with an SVM approach reaching roughly 80% and its own earlier GLM reaching 81% on a selected 60-second segment, and argues that the CNN matches these without handcrafted features — but those benchmarks come from different evaluation setups, so the comparison is indicative rather than like-for-like.

Where the evidence points next

If the goal is a subject-independent marker of hypnotic susceptibility from minimally preprocessed EEG, this paper's evidence points to anterior cortex and to window aggregation, not to the posterior regions that dominate a naive random split. The frontal region gave the best single-window LOSO accuracy at 0.646, and the prefrontal region gave the best ensemble result at 0.738 with AUC 0.761. The occipital region, despite 90.3% of windows passing screening, is the clearest warning: its near-chance LOSO AUC of 0.429 shows that a high random-split score can reflect individual EEG signatures rather than a group effect.

The most useful threshold in the paper is therefore not a single accuracy figure but the gap between the two evaluation regimes. Any future claim about EEG-based hypnotic-susceptibility classification should report subject-independent performance per region, because the region ranking and the apparent skill both change when leakage is removed.

Sources used in this answer