💻 computer science

AutoML-XAI convolutional neural network for hypnotic susceptibility classification from segmented multichannel EEG

This paper proposes an interpretable AutoML-driven CNN framework that classifies hypnotic susceptibility from segmented multichannel EEG data, achieving perfect accuracy in its best fold and identifying alpha and gamma frequency bands as the most critical features through XAI techniques.

Original authors: Ali Walid Daher

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Walid Daher

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to tell the difference between two types of people: those who are easily "hypnotized" (highly susceptible) and those who are not (lowly susceptible). Usually, scientists look at brain waves (EEG) to do this, but it's like trying to find a needle in a haystack because the data is messy, the brain signals are complex, and there isn't much of it to work with.

This paper presents a new, smart way to solve this puzzle using a team of three digital tools: AutoML, CNNs, and XAI. Here is how they work together, explained simply:

1. The Setup: Cutting the Brain Waves into "Snippets"

The researchers started with recordings of 16 different brain sensors. Instead of looking at the whole recording as one giant block, they treated it like a long loaf of bread.

The Slicing: They sliced each recording into 10 small, non-overlapping pieces called "snippets."
The Format: They organized these slices into a grid (a matrix) that looks like a spreadsheet with 5 rows (representing 5 different speed ranges of brain waves: Delta, Theta, Alpha, Beta, and Gamma) and 1,600 columns (representing time).
The Golden Rule: To make sure the computer didn't cheat, they kept all 10 slices from the same person together. If a person's slices went into the "training" group, none of them went into the "testing" group. This prevents the computer from just memorizing the person's name instead of learning the pattern.

2. The Team of Three

The Architect (AutoML)

Think of AutoML as a super-robot architect. Building a neural network (a type of AI brain) usually requires a human to guess how many layers it needs, how big the filters should be, and how fast it should learn.

What it did: The robot architect tried thousands of different building designs automatically. It asked, "What if I add a layer here? What if I change the learning speed?" It picked the single best design that worked well on the training data.

The Detective (CNN)

The Convolutional Neural Network (CNN) is the actual detective that looks at the brain-wave grid.

What it did: Once the architect picked the best design, the detective started scanning the 5x1,600 grids. It looked for patterns in the brain waves that signaled "High Susceptibility" or "Low Susceptibility."

The Translator (XAI)

This is the most important part. Usually, AI is a "black box"—you put data in, and it gives an answer, but you don't know why. XAI (Explainable AI) is like a translator that opens the black box and explains the detective's reasoning.

What it did: The researchers used tools called SHAP and LIME to ask the detective, "Which rows of your spreadsheet were you looking at to make that decision?"
The Discovery: The translator revealed that the detective was ignoring some rows and focusing heavily on just two: Alpha and Gamma waves.

3. The Results: Two Different Stories

The paper tested this system in two ways, and the results were interesting:

The "Perfect" Moment: In one specific test run, the researchers took the detective's findings (that Alpha and Gamma were important) and gave those specific rows extra weight. In this one lucky scenario, the AI got a perfect score (100% accuracy). It correctly identified every single person in that test group.
The "Realistic" Average: However, when they tested the system across many different groups (to make sure it wasn't just luck), the perfect score disappeared. The average accuracy dropped to about 67%.
- The Takeaway: The "perfect" score shows the method can work, but the lower average score shows that the system is still learning and needs more data to be reliable.

4. What Did They Learn?

The study found that when you use this smart, automated approach, the Alpha and Gamma frequency bands of the brain waves seem to hold the most clues for telling if someone is hypnotizable.

Summary Analogy

Imagine you are trying to identify a specific type of bird by looking at a 100-page photo album of its feathers.

Old way: You manually pick the best pages and guess the rules.
This paper's way:
- AutoML is a robot that tries every possible way to organize the album to find the best view.
- The CNN is the bird-watcher who studies the album.
- XAI is the bird-watcher's notebook where they write down, "I didn't look at pages 1–50; I only looked at pages 51–60 because that's where the unique feather patterns were."
- The Result: The bird-watcher found the unique patterns (Alpha and Gamma), but because there are only a few photo albums to study, they aren't 100% sure they can identify the bird in every new album yet.

The Bottom Line: The paper proves that combining automated design (AutoML) with explainable tools (XAI) is a promising way to study hypnotic susceptibility using brain waves, but it warns that we need more data before we can trust it completely.

Technical Summary: AutoML-XAI CNN for Hypnotic Susceptibility Classification

Problem Statement
The classification of hypnotic susceptibility using electroencephalography (EEG) remains an underexplored and methodologically inconsistent field compared to other EEG tasks such as seizure detection or sleep staging. Existing studies often rely on classical feature extraction (e.g., differential entropy) or specific machine learning models without fully integrating automated architecture search with explainable artificial intelligence (XAI). Furthermore, the field suffers from limited data availability, leading to challenges in model generalization and the frequent use of synthetic data augmentation, which may introduce neurophysiological assumptions that do not hold true. There is a specific gap in combining AutoML-driven Convolutional Neural Network (CNN) optimization with XAI-guided frequency band interpretation to create an interpretable, data-efficient model for hypnotic susceptibility.

Methodology
The study proposes a framework utilizing the public nplresearch/hypnotic-state dataset, which contains 16-channel EEG recordings from subjects categorized as having high or low hypnotic susceptibility.

Data Representation and Segmentation:
- Preprocessing: Each 16-channel recording is filtered into five standard frequency bands: Delta (1–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), Beta (13–30 Hz), and Gamma (30–45 Hz).
- Resampling and Formatting: Signals are resampled to 1,000 time points per channel. The data is then split into ten non-overlapping "snippets" per subject.
- Input Matrix: Each snippet is formatted as a $5 \times 1600$ matrix. The 5 rows represent the frequency bands, and the 1,600 columns preserve the ordered time-series structure (16 channels $\times$ 100 time points per channel).
- Non-Synthetic Expansion: Unlike many deep learning approaches, this method does not generate synthetic data or use cross-subject recombination. It relies solely on observed EEG data, increasing sample size through segmentation while maintaining signal integrity.
Validation Strategy:
- To prevent data leakage, the study employs a grouped K-fold cross-validation strategy. All ten snippets derived from a single subject are kept together within either the training or testing set for a given fold. This ensures that the model is evaluated on unseen subjects rather than unseen snippets of seen subjects.
AutoML Optimization:
- The CNN architecture and hyperparameters are optimized using Optuna. The search space includes the number of convolutional layers, filter counts, kernel dimensions, batch normalization, dropout rates, dense layer units, learning rates, and batch sizes.
- The objective function maximizes the mean validation balanced accuracy across the grouped folds.
Explainable AI (XAI) and Model Refinement:
- After training the best-performing full-band CNN, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are applied to identify the most relevant frequency bands.
- Based on these XAI insights, three model variants are constructed and tested:
  1. A CNN using only the top two XAI-ranked frequency bands.
  2. A full-band CNN with row weights derived from XAI importance scores.
  3. A stress-test variant where the top two bands are multiplied by a factor of ten.
- Subject-level predictions are aggregated by averaging the probabilities of the ten snippets for each subject.

Key Results

Architecture: The Optuna search selected a CNN with four convolutional layers, batch normalization, a dropout rate of ~0.19, a dense layer of 256 units, a learning rate of $5.53 \times 10^{-5}$ , and a batch size of 2.
Performance Metrics:
- Best-Fold Performance: The XAI-row-weighted CNN achieved a perfect subject-level accuracy, balanced accuracy, F1 score, and AUC of 1.00 in a single best fold.
- Cross-Fold Performance: The Top-Two-Band XAI model (using only Alpha and Gamma bands) demonstrated the most robust mean performance across all folds, achieving a mean subject-level accuracy and balanced accuracy of 0.667, and a mean AUC of 0.689.
Frequency Band Importance: The combined SHAP and LIME analysis identified Alpha and Gamma bands as the two most critical features for classification, followed by Beta, Theta, and Delta.
- SHAP rankings highlighted Gamma-band features (specifically around the P3 channel) as dominant.
- LIME distributed relevance more broadly but still supported the high importance of Alpha and Gamma.

Significance and Claims
The paper claims that this work represents the first integration of AutoML-based CNN optimization with XAI-guided frequency band selection and reweighting specifically for hypnotic susceptibility classification.

Interpretability: The study moves beyond "black-box" deep learning by using XAI not just for post-hoc explanation but to actively guide model refinement (band selection and weighting).
Data Efficiency: By utilizing a conservative, non-synthetic segmentation strategy with grouped validation, the authors demonstrate that high-quality models can be developed without relying on potentially biased data augmentation techniques.
Feasibility: The results suggest that AutoML and XAI can be effectively combined to produce an interpretable EEG model for hypnosis. The identification of Alpha and Gamma as key discriminators aligns partially with existing literature on hypnotic depth and suggestibility, offering a new perspective through the lens of deep learning feature extraction.

Limitations and Future Outlook
The authors maintain a modest tone regarding the generalizability of their findings. They explicitly state that the perfect best-fold result (1.00 accuracy) should be viewed as a promising exploratory outcome rather than conclusive evidence of generalization, given the lower mean performance across folds. The paper emphasizes the critical need for larger, independent validation datasets to confirm these findings and acknowledges that the current results are specific to the pipeline and dataset used.