ADHDeepNet: CNN Model for ADHD EEG Diagnosis

• ADHDeepNet is a deep learning model that processes raw EEG to automatically distinguish ADHD from healthy controls using a multi-branch CNN architecture.
• It integrates temporal convolution, depthwise spatial filtering, and adaptive channel-wise attention to learn discriminative features directly from unprocessed signals.
• Robust cross-validation and targeted AGN augmentation yield high sensitivity and accuracy, while explainability methods like t-SNE and filter analysis enhance clinical insight.

Searching arXiv for the specified paper to ground the article in the source. ADHDeepNet is a deep learning model for automatic diagnosis of Attention Deficit Hyperactivity Disorder (ADHD) from electroencephalogram (EEG) recordings. It is presented as a CNN-based architecture for ADHD versus healthy control (HC) classification that operates directly on raw segmented EEG, combining temporal-spatial processing, an Inception/Xception-inspired module termed InXception, adaptive channel-wise attention via Squeeze-and-Excitation (SE) blocks, additive Gaussian noise (AGN) augmentation, nested cross-subject validation, and post hoc explainability analyses based on learned filters, spatial weights, and t-distributed Stochastic Neighbor Embedding (t-SNE) (Amini et al., 10 Sep 2025).

1. Clinical objective and problem formulation

ADHDeepNet is designed to support automatic diagnosis of ADHD using EEG. The underlying clinical motivation is that ADHD diagnosis is traditionally based on behavioral interviews and clinical criteria and is described as labor-intensive, time-consuming, and potentially ambiguous in routine practice. The model is therefore framed as an EEG-based deep learning system for more objective, fast, and accurate ADHD screening or diagnosis (Amini et al., 10 Sep 2025).

The paper positions the model against several limitations in earlier ADHD/EEG pipelines. First, many prior methods rely on handcrafted features, including linear, nonlinear, spectral, connectivity, or image-transformed representations. Second, previous CNN-based approaches often convert EEG into spectrograms, RGB images, or connectivity matrices rather than processing the signal directly. Third, robustness and overfitting are treated as methodological concerns, motivating nested cross-validation and augmentation. Fourth, interpretability is emphasized through analysis of temporal filters, spatial weights, and latent representations. In this formulation, ADHDeepNet is intended to learn discriminative structure directly from raw multichannel EEG rather than from manually engineered descriptors.

A central operational distinction is that inference is performed at the segment level but clinical classification is reported at the subject level by aggregating segment-wise predicted probabilities. This preserves the use of short fixed-length windows while keeping the final decision aligned with subject diagnosis.

2. Data source, acquisition setting, and signal representation

The model was evaluated on a dataset of 121 children, comprising 61 ADHD cases and 60 healthy controls. The children were boys and girls aged 7–12 years. ADHD diagnosis was made according to DSM-IV criteria by an experienced psychiatrist. Healthy controls were required to have no severe head injury, no epilepsy, no drug abuse, no psychiatric disorder, and no high-risk behavior history. EEG was recorded at the Psychology and Psychiatry Research Center, Roozbeh Hospital, Tehran, Iran (Amini et al., 10 Sep 2025).

The recordings were obtained during a visual counting task rather than during simple resting-state acquisition. Children were shown figures or pictures and were asked to enumerate them. This is described as a continuous cognitive or visual attention task. A plausible implication is that the classification problem is conditioned not only on intrinsic EEG differences but also on task-evoked differences in attentional processing, although no additional task-specific modeling is reported.

EEG acquisition used 19 channels at a sampling rate of 128 Hz. The listed electrodes are

$$\{ \text{Fz, Cz, Pz, C3, T3, C4, T4, Fp1, Fp2, F3, F4, F7, F8, P3, P4, T5, T6, O1, O2} \}.$$

The amplifier or hardware model, reference montage, impedance criteria, and exact recording duration per subject were not specified. The recording length varied slightly across subjects.

Each subject’s EEG was represented as

$X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$

where $C=19$ is the number of channels and $p_j$ is the total number of time points for subject $j$. Signals were segmented into non-overlapping 4-second epochs. Since $f_s=128$ Hz, each segment contains

$T = 4 \times 128 = 512$

time samples, yielding segment tensors

$$X_{(j)}^{s} \in \mathbb{R}^{E \times T}, \quad E=19,\; T=512.$$

Across all subjects, the ADHD group contributed 2330 segments and the HC group 1843 segments, for a total of 4173 segments. Each segment inherited the subject-level label: $$Y_{(j)}^s = \begin{cases} (1,0) & \text{if } y_{(j)}^s = \text{ADHD} \ (0,1) & \text{if } y_{(j)}^s = \text{HC}. \end{cases}$$

The paper explicitly states that no additional preprocessing was applied beyond segmentation. It reports that no extra filtering, artifact rejection, ICA, re-referencing, band-pass filtering, normalization, or handcrafted feature extraction was used at the modeling stage. No channel reduction was performed; all 19 channels were retained.

3. Architecture and temporal-spatial processing pipeline

ADHDeepNet is described as an enhanced EEGNet with a reported total of 228,642 parameters. It preserves EEGNet’s early temporal and spatial inductive bias and modifies the deeper processing stages by adding the InXception module, SE attention blocks, a later separable convolution, and global average pooling (GAP) instead of flattening (Amini et al., 10 Sep 2025).

The architecture begins with a temporal convolution layer operating on raw EEG segments of conceptual shape

$$X \in \mathbb{R}^{E \times T}, \quad E=19,\; T=512.$$

In the explainability analysis, this first temporal layer is described as having 64 kernels of size $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$0. The interpretation given in the paper is that these kernels span time only and learn temporal filters analogous to digital band-pass-like filters. Batch normalization follows this layer.

A depthwise convolution is then used to learn spatial features across EEG channels. This stage is described as isolating spatial information by applying a kernel individually relative to the 19 channels. Batch normalization and ELU activation follow. The temporal-spatial decomposition therefore proceeds in the EEGNet style: temporal filtering first, then spatial filtering conditioned on temporal structure.

After this block, the network applies average pooling with parameter $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$1, reducing output volume by 50%, followed by dropout. The core novelty then appears in the InXception module, which uses four parallel streams whose outputs are concatenated. Two of the branches are described as combining outputs via point-wise convolution followed by 2D separable convolutions with filter dimensions $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$2 and $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$3. The full specification of all four branches is not reported, but the intended role is explicit: capture multi-scale features and richer spatiotemporal structure while using separable convolutions to reduce the number of parameters relative to standard convolutions.

The first SE block follows hierarchical feature extraction. Given a feature map $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$4, channel descriptors are obtained by global average pooling: $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$5 A two-layer MLP models inter-channel dependencies: $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$6 and the resulting channel importance weights rescale the feature maps: $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$7 This implements adaptive channel-wise recalibration.

A separable 2D convolution with kernel size $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$8, followed by batch normalization and ELU, acts as a further refinement stage. A second SE block is then applied, so attention is used twice: once after multi-scale feature extraction and once after later refinement. The classification head consists of dropout, GAP, a fully connected layer with two outputs, and softmax: $X_{(j)}^{c,i} \in \mathbb{R}^{C \times p_j},$9

This sequence can be summarized as temporal convolution, depthwise spatial convolution, pooling and dropout, InXception multi-branch extraction, SE attention, separable convolution refinement, second SE attention, and GAP-based classification.

4. Optimization, augmentation, and evaluation protocol

The model is trained with a softmax-based negative log-likelihood or cross-entropy objective: $C=19$0 Hyperparameters are selected using Bayesian Optimization (BO) within an inner 2-fold cross-subject validation loop. The tuned hyperparameters are norm-rate, learning rate, optimizer type, dropout rate, and batch size. The BO objective is negative validation accuracy,

$C=19$1

with a Gaussian Process surrogate and an acquisition function

$C=19$2

For each outer fold, BO is run for 100 iterations (Amini et al., 10 Sep 2025).

The evaluation protocol is a nested $C=19$3-fold cross-subject validation procedure. The outer loop is 10-fold cross-subject cross-validation. The inner loop is 2-fold cross-subject cross-validation for hyperparameter optimization. All segments from a subject are kept in the same fold, no overlapping windows are used, and augmentation is applied only to training data. These choices are presented as safeguards against leakage and correlation-induced overfitting.

The training methodology is two-stage. In Stage 1, each outer fold uses inner-loop BO to choose hyperparameters. In Stage 2, augmentation is applied only to the training set, after which the model is retrained and evaluated on the untouched outer test fold. AGN augmentation is defined as

$C=19$4

where the noise is Gaussian with zero mean and standard deviation $C=19$5. The tested ranges are

$C=19$6

The paper reports both single $C=19$7 combinations and pairwise combinations of augmentation settings, summarized as

$C=19$8

Final subject prediction aggregates segment-level probabilities: $C=19$9 Evaluation uses accuracy, precision, recall, and $p_j$0, with

$p_j$1

Because false negatives are considered particularly undesirable in ADHD screening, the paper emphasizes $p_j$2.

5. Reported performance and ablation findings

Without augmentation, nested $p_j$3-fold cross-subject validation yielded the following results for ADHDeepNet and EEGNet (Amini et al., 10 Sep 2025):

Model/setting	Accuracy	F2
ADHDeepNet sample-level	$p_j$4	$p_j$5
ADHDeepNet subject-level	$p_j$6	$p_j$7
EEGNet sample-level	$p_j$8	$p_j$9
EEGNet subject-level	$j$0	$j$1

Fold-wise subject accuracy for ADHDeepNet without augmentation was reported as 0.9167, 0.8462, 1.0000, 0.9167, 0.9167, 0.9167, 0.9167, 1.0000, 0.9167, and 0.9167 across folds 1–10. This indicates reasonably stable cross-subject performance with one weaker fold.

With augmentation, the optimal ADHDeepNet setting produced sample accuracy $j$2, subject accuracy $j$3, sample $j$4 $j$5, and subject $j$6 $j$7. Under the same table, EEGNet with optimal augmentation achieved sample accuracy $j$8, subject accuracy $j$9, sample $f_s=128$0 $f_s=128$1, and subject $f_s=128$2 $f_s=128$3. The best reported augmentation setting for ADHDeepNet was $f_s=128$4, $f_s=128$5, while the worst reported setting was $f_s=128$6, $f_s=128$7. This suggests that small perturbations were beneficial, whereas aggressive noise degraded class-discriminative structure.

The paper explicitly states that ADHDeepNet achieved 100% sensitivity and 99.17% accuracy in classifying ADHD/HC subjects, and also reports AUC $f_s=128$8. It further reports a low error rate of 0.41% in sample accuracy between training and validation sets as evidence that overfitting was controlled.

An ablation study evaluated the InXception module and the SE blocks. Without augmentation, the full ADHDeepNet achieved sample accuracy 0.9012, subject accuracy 0.9256, sample $f_s=128$9 0.9372, and subject $T = 4 \times 128 = 512$0 0.9558. A variant with Inception and without SENet achieved sample accuracy 0.8708 and subject accuracy 0.8760, while a variant with SENet and without Inception achieved sample accuracy 0.8789 and subject accuracy 0.8842. Both modules therefore contributed to performance, and the best results required their combination.

6. Explainability, neurophysiological interpretation, and limitations

The explainability strategy rests on two analyses: inspection of learned temporal and depthwise convolution weights, and t-SNE visualization of hidden activations from different layers (Amini et al., 10 Sep 2025).

For the 8th fold, the first temporal convolution layer was analyzed as a bank of 64 filters of shape $T = 4 \times 128 = 512$1. The coefficients $T = 4 \times 128 = 512$2 of filter $T = 4 \times 128 = 512$3 were treated as digital filter coefficients, with frequency response

$T = 4 \times 128 = 512$4

Average amplitude in canonical EEG bands was then computed as

$T = 4 \times 128 = 512$5

The analyzed bands were $T = 4 \times 128 = 512$6, $T = 4 \times 128 = 512$7, $T = 4 \times 128 = 512$8, and $T = 4 \times 128 = 512$9. The paper reports that some filters showed a high theta-to-beta ratio, which the authors interpret as aligned with known ADHD EEG biomarkers.

For spatial interpretation, the weights of the depthwise convolution associated with the temporal filters were normalized to $$X_{(j)}^{s} \in \mathbb{R}^{E \times T}, \quad E=19,\; T=512.$$0 and mapped onto scalp topographies. The topographical maps highlighted the frontal region as especially active or informative. The paper does not provide a ranked channel list, but the frontal area is the main regional emphasis.

The t-SNE analysis projected activation outputs from three different layers into two dimensions for the 8th fold. Early-layer embeddings showed weaker separation between ADHD and HC, while later-layer embeddings showed clearer class discrimination. This suggests that deeper layers progressively transformed raw EEG into more class-separable representations.

The same paper also defines the principal limitations of the method. Evaluation was performed on a single dataset and no external validation was presented. Although 4173 segments were available, they came from only 121 subjects, so the effective independent sample size remained limited. Important implementation details were not reported, including final optimizer choice, final learning rate, final batch size, number of epochs, exact dropout value, exact branch structure, exact tensor shapes, software framework, hardware, training time, random seed control, and code availability. Preprocessing transparency is also limited because no discussion is provided for line noise, artifact contamination, baseline drifts, or referencing. Finally, the interpretability analysis is limited to filter spectra, scalp maps, and t-SNE; no saliency maps, Grad-CAM, CAM, SHAP, or integrated gradients are reported.

These constraints imply that the headline subject-level results should be interpreted within the scope of the reported cohort and protocol. The paper nonetheless presents ADHDeepNet as a compact, EEG-specialized, interpretable raw-signal deep network that advances beyond EEGNet for ADHD/HC classification on the evaluated dataset.