Explainable electroencephalography-based attention-deficit/hyperactivity disorder detection model with a combination of ternary pattern and twin wavelet transform

Table 1 Details of the dataset

Type	Female	Male	Total	Age (mean ± SD)
ADHD	49	88	137	7.86 ± 3.28
Control	82	68	150	9.18 ± 2.34

ADHD: Attention-deficit/hyperactivity disorder.

Table 2 Details of various components of the proposed model

Layer	Method(s)	Parameter(s)	Output(s)
Feature extraction	Twin wavelet transformation	Filter: Sym4; number of levels: 7	13 wavelet bands
	Combination ternary pattern	Kernel: Ternary; block length: 5; pattern: Combination	Feature vector length: 486
	Statistical features	20 statistical moments	Feature vector length: 40
	Concatenate features	Textural and statistical features	Feature vector length: 526
	Concatenate feature vectors	Merge 14 feature vectors	Feature vector length: 7364
Feature selection	Neighborhood component analysis	Select the top informative features	Selected feature vector length: 263
Classification	k-Nearest neighbors	k:1; distance: L2norm, voting: No	20 prediction vectors
Post-processing	Iterative majority voting	Loop range: 3 to 20; function: Mode	18 voted vectors
	Greedy algorithm	Maximum accuracy	Best result

Table 3 Channel-wise classification performance based on k-nearest neighbors using 10-fold cross-validation, %

Channel	Acc	Sen	Spe	Channel	Acc	Sen	Spe
1	98.92	98.32	99.25	11	98.96	98.09	99.43
2	98.55	97.79	98.97	12	98.72	97.79	99.23
3	98.86	98.17	99.25	13	98.91	98.32	99.23
4	98.53	97.79	98.93	14	98.77	97.79	99.31
5	99.08	98.51	99.39	15	98.80	98.09	99.18
6	98.96	98.32	99.31	16	98.82	98.06	99.25
7	97.99	96.65	98.72	17	98.93	98.48	99.18
8	99.12	98.74	99.33	18	98.69	98.17	98.97
9	99.08	98.59	99.35	19	99.04	98.63	99.27
10	98.91	98.78	98.97	20	98.82	98.06	99.25

Acc: Accuracy the overall percentage of correctly classified samples (both true positives and true negatives) out of all test samples; Sen: Sensitivity (also called Recall or True Positive Rate) is the percentage of actual positive cases correctly identified by the classifier; Spe: Specificity (also called True Negative Rate), the percentage of actual negative cases correctly identified as negative.

Table 4 Overall model performance with incremental iterative mode function-based majority voting, %

Iteration	Acc	Sen	Spe	Iteration	Acc	Sen	Spe
1	99.70	99.39	99.87	10	99.96	99.92	99.98
2	99.82	99.92	99.77	11	99.92	99.81	99.98
3	99.89	99.77	99.96	12	99.95	99.89	99.98
4	99.96	100	99.94	13	99.95	99.89	99.98
5	99.96	99.89	100	14	99.95	99.89	99.98
6	99.97	99.96	99.98	15	99.95	99.85	100
7	99.93	99.85	99.98	16	99.93	99.85	99.98
8	99.97	100	99.96	17	99.93	99.81	100

Acc: Accuracy, the overall percentage of correctly classified samples (both true positives and true negatives) out of all test samples; Sen: Sensitivity (also called Recall or True Positive Rate) is the percentage of actual positive cases correctly identified by the classifier; Spe: Specificity (also called True Negative Rate), the percentage of actual negative cases correctly identified as negative.

Table 5 Comparison of results, %

Ref.	Method(s)	Classifier	Subjects	Samples	Channels	Split ratio	Results
Nouri et al[50], 2023	Layer-wise relevance propagation, CNN	Softmax	31 ADHD; 30 controls	656	19	5-fold CV	Acc: 92.45; Sen: 93.06; Spe: 98.10
Chen et al[24], 2019	CNN	Softmax	50 ADHD; 51 controls	4545	32	90:10	Acc: 94.67
Sharma et al[51], 2023	MEMD, GA, MEWT, multivariate empirical-basis decomposition approaches, NCA	ANN	61 ADHD; 60 controls	7983	19	5-fold CV	Acc: 96.16; F1: 96.32; MCC: 0.92
Cura et al[52], 2023	Intrinsic time-scale decomposition	Bagged tree	15 ADHD; 18 controls	198	12	10-fold CV	Acc: 99.46; Sen: 99.47; Spe: 99.47
Barua et al[53], 2022	Ternary motif pattern, TQWT, NCA	kNN	61 ADHD 60 Controls	4173	14	10-fold CV	Acc: 95.57; GM:95.18
Tor et al[54], 2021	EMD, DWT	kNN	45 ADHD; 62 ADHD + CD; CD 16	5000	12	10-fold CV	Acc: 97.88; Sen: 96.68; Spe: 100
Tosun[55], 2021	LSTM, PSD	SVM	8 ADHD; 8 controls	4352	16	80:20	Acc:92.15; Sen: 90.95; Spe: 93.43
Moghaddari et al[56], 2020	Deep CNN	Softmax	31 ADHD; 30 controls	328	19	5-fold CV	Acc: 98.48; Rec:98.48; F1:98.49; Pre: 98.51
Kaur et al[57], 2020	Phase space reconstruction	SVM	47 ADHD; 50 controls	Not specified	19	69:31	Acc: 93.30; Sen: 100; Spe: 86.70
Tanko et al[58], 2022	8-pointed star pattern learning network	kNN	61 ADHD; 60 controls	Not specified	19	10-fold CV	Acc: 97.19; Rec: 97.12; Pre: 97.18; F1: 97.15
García-Ponsoda et al[59], 2024	Independent component analysis process	XGBoost	61 ADHD; 60 controls	128 samples per second	19	5-fold CV	Acc: 86.10
Mercado-Aguirre et al[60], 2025	Ridge Regression, Independent Component Analysis	SVM	22 ADHD; 25 controls	800 ms (100 samples)	6	5-fold CV	Acc: 86.36; Sen: 95.45
Colonnese et al[61], 2025	Hyperdimensional Computing with a spatio-temporal encoder	ADHDC (HDC)	37 ADHD; 42 controls	7168	14	75:25	Acc: 88.90; F1: 87.50; Rec: 90.40
Cai et al[62], 2025	Phase space reconstruction	kNN	61 ADHD; 60 controls	Not specified	19	80:20	Acc: 78.27; Sen: 80.62; Spe: 75.63
Alhussen et al[63], 2025	Discrete Cosine Transform- Independent Component Analysis, rhinofish optimization, AttentionNet	Softmax	(1) 17 ADHD; 17 controls; (2) 51 ADHD; 52 controls	Not specified	25	5-fold CV	(1) Acc: 97.89; (2) Acc: 98.52
Our study	Twin wavelet transformation, combination ternary pattern, NCA, iterative majority voting	kNN	137 ADHD; 150 controls	7399	20	10-fold CV	Acc: 99.97; Sen: 99.96; Spe: 99.98

CNN: Convolutional neural network; ADHD: Attention-deficit/hyperactivity disorder; CV: Cross-validation; MEMD: Multivariate empirical mode decomposition; GA: Genetic algorithm; MEWT: Multivariate empirical wavelet transform; NCA: Neighborhood component analysis; MCC: Matthews correlation coefficient; TQWT: Tunable Q-factor wavelet transform; EMD: Empirical mode decomposition; DWT: Discrete wavelet transform; CD: Conduct disorder; LSTM: Long short-term memory; PSD: Power spectral density; SVM: Support vector machine; HDC: Hyperdimensional computing; Acc: Accuracy; Sen: Sensitivity; Spe: Specificity.