RatioWaveNet

RatioWaveNet is a custom deep learning architecture designed for the classification of EEG signals in Motor Imagery (MI) tasks. Inspired by existing models like EEGNet and ATCNet, RatioWaveNet introduces a lightweight CNN framework enhanced with residual connections, dropout, and optimized temporal feature extraction.

Authors : Giuseppe Bonomo, Marco Siino, Rosario Sorbello

University of Palermo, Italia

---

In addition to the proposed RatioWaveNet model, the repository includes implementations of several other well-known EEG classification architectures in the models.py file, which can be used as baselines for comparison with RatioWaveNet. These include:

• ATCNet:paper, original code
• EEGNet:paper, original code
• EEG-TCNet:paper, original code
• MBEEG_SENet:paper
• ShallowConvNet:paper, original code

The following table summarizes the classification performance of RatioWaveNet and the other reproduced models, based on the experimental setup defined in the notebook for each model and dataset.

Model	#params	BCI 4-2a Accuracy	BCI 4-2a Kappa	BCI 4-2b Accuracy	BCI 4-2b Kappa	HGD Accuracy	HGD Kappa
RatioWaveNet	113,732	82.14	76.20	97.69	80.60	92.81	90.40
ATCNet	113,732	81.10	79.73	89.41	78.80	92.05	89.40
EEGTCNet	4,096	77.97	70.63	83.69	67.31	87.80	83.73
MBEEG_SENet	10,170	79.98	73.30	86.53	73.02	90.13	86.84
ShallowConvNet	47,310	80.52	74.02	86.02	72.38	87.00	82.67
EEGNet	2,548	77.68	70.24	86.08	72.13	88.25	84.33

Disclaimer

• The results reported for BCI 4-2a, BCI 4-2b and HGD datasets were not recomputed by us and are directly extracted from the original papers.
• HGD (High Gamma Dataset): Refers to physically executed movements (executed movements), not motor imagery (motor imagery).

---

Comparative preprocessing

The following table presents a comparative analysis of different deep learning models with and without the application of the RDWT (Rational Dilated Wavelet Transform) preprocessing technique. The evaluation covers three benchmark EEG motor imagery datasets: BCI Competition IV-2a, BCI Competition IV-2b, and the High-Gamma Dataset (HGD). The aim is to assess the impact of RDWT on classification performance (accuracy and Cohen’s kappa score).

In particular, the results highlight the performance improvements achieved by RatioWaveNet when combined with the RDWT preprocessing, compared to both its baseline (no preprocessing) and other well-established architectures.

Model	Preprocessing	BCI 2a Acc.	BCI 2a κ	BCI 2b Acc.	BCI 2b κ	HGD Acc.	HGD κ
RatioWaveNet	None	79.36	72.50	97.00	69.80	87.45	83.30
	RDWT	82.14	76.20	97.69	80.60	92.81	90.40
ATCNet	None	79.71	72.90	96.90	63.30	88.88	85.20
	RDWT	79.51	72.70	96.74	61.90	88.26	84.30
EEGTCNet	None	64.35	52.50	95.81	58.90	86.60	82.10
	RDWT	68.79	58.40	96.09	66.60	87.14	82.90
MBEEG_SENet	None	70.49	60.60	96.95	73.80	90.58	87.40
	RDWT	72.72	63.60	96.28	63.50	90.26	87.00
ShallowConvNet	None	65.74	54.30	96.13	60.70	87.05	82.70
	RDWT	66.32	55.10	95.94	62.30	87.27	87.27
EEGNet	None	70.79	61.10	95.85	59.60	87.32	83.10
	RDWT	70.10	60.10	96.06	64.00	88.08	84.10

Note

• The recomputed results for these datasets (including accuracy/kappa scores) are available in their respective dataset folders.
• Unlike the previous table, the results reported here for the HGD and BCI IV-2b datasets include an enhanced preprocessing pipeline, which incorporates data augmentation and class balancing techniques. These strategies were employed to address class imbalance and improve the generalization capabilities of the models.
• These values were obtained using our implementation and preprocessing pipeline. Minor deviations from the original papers are expected.

---

Multi-seed robustness (multiple runs) — RatioWaveNet vs ATCNet

To assess robustness to training stochasticity (i.e., to ensure that the observed improvements are not driven by a favorable initialization), we evaluate RatioWaveNet and ATCNet across multiple independent random seeds on three benchmarks: BCI Competition IV-2a, BCI Competition IV-2b, and the High-Gamma Dataset (HGD).
We report subject-wise test accuracy (%) for each run, together with Avg Acc (mean accuracy across subjects) and Avg κ (mean Cohen’s κ across subjects). The final Mean row summarizes results across runs (5 seeds for BCI IV-2a and IV-2b; 3 seeds for HGD).

BCI IV-2a — Detailed test accuracy across 5 seeds

Model	Seed	S1	S2	S3	S4	S5	S6	S7	S8	S9	Avg Acc	Avg κ
RatioWaveNet	1	86.46	68.75	93.06	84.03	74.65	70.83	92.01	80.21	89.24	82.14	0.762
	2	86.11	62.85	92.36	82.29	71.18	69.44	93.75	81.25	87.50	80.75	0.743
	3	85.76	67.71	91.32	83.33	64.93	69.44	86.81	79.86	85.42	79.40	0.725
	4	82.64	67.01	94.44	84.03	69.79	70.49	90.28	77.08	86.46	80.25	0.737
	5	85.42	63.19	89.93	76.39	71.53	76.04	91.32	79.86	89.93	80.40	0.739
	Mean	85.28	65.90	92.22	82.01	70.42	71.25	90.83	79.65	87.71	80.59	0.741
ATCNet	1	86.46	70.49	92.71	78.12	71.88	66.67	92.36	77.08	87.15	80.32	0.738
	2	85.76	65.97	93.06	84.72	72.22	67.36	90.28	75.35	86.81	80.17	0.736
	3	85.42	67.01	94.10	76.04	71.53	70.14	91.67	76.74	87.15	79.98	0.733
	4	85.42	63.19	89.58	77.43	64.93	70.83	93.06	79.51	90.97	79.44	0.726
	5	86.11	63.54	88.54	80.21	71.53	68.06	93.06	80.21	86.11	79.71	0.729
	Mean	85.83	66.04	91.60	79.30	70.42	68.61	92.70	77.78	87.64	79.92	0.732

BCI IV-2b — Detailed test accuracy across 5 seeds

Model	Seed	S1	S2	S3	S4	S5	S6	S7	S8	S9	Avg Acc	Avg κ
RatioWaveNet	1	98.94	99.10	97.46	91.11	92.31	100.00	98.89	99.28	99.13	97.36	0.656
	2	98.94	97.30	98.91	97.78	91.67	99.53	98.52	99.64	99.13	97.93	0.732
	3	98.94	98.20	98.91	91.11	92.95	100.00	98.15	98.19	99.13	97.29	0.665
	4	98.94	98.20	97.83	88.89	93.59	100.00	98.89	99.28	98.70	97.14	0.617
	5	98.94	98.20	98.19	88.89	91.03	100.00	98.15	99.64	99.57	96.95	0.649
	Mean	98.94	98.20	98.26	91.56	92.31	99.81	98.52	99.21	99.13	97.34	0.664
ATCNet	1	100.00	98.20	99.28	82.22	92.31	100.00	98.89	98.55	98.70	96.46	0.676
	2	98.94	98.20	99.64	86.67	91.67	99.53	97.78	98.55	97.40	96.49	0.569
	3	98.94	97.30	98.91	86.67	94.23	99.06	100.00	99.28	96.10	96.72	0.631
	4	98.23	99.10	98.19	86.67	90.38	99.53	96.30	98.91	98.27	96.17	0.521
	5	99.29	98.20	98.91	86.67	92.31	100.00	96.67	98.91	99.57	96.72	0.637
	Mean	99.08	98.20	98.99	85.78	92.18	99.62	97.93	98.84	97.81	96.51	0.607

Takeaway (2b): The task is near-ceiling for both methods, but RatioWaveNet maintains a consistent edge in Avg Acc and, more importantly, a higher Avg κ, which is informative under saturation.

HGD — Detailed test accuracy across 3 seeds

Model	Seed	S01	S02	S03	S04	S05	S06	S07	S08	S09	S10	S11	S12	S13	S14	Avg Acc	Avg κ
RatioWaveNet	1	87.50	92.50	98.12	97.50	92.50	95.00	91.82	90.00	93.75	88.12	85.00	90.62	89.38	79.38	90.80	0.877
	2	86.88	95.62	98.12	98.12	93.12	90.00	89.31	91.88	95.00	90.00	80.62	94.38	89.38	76.88	90.66	0.876
	3	90.62	90.00	98.75	97.50	95.00	92.50	91.19	91.25	95.00	93.75	81.88	90.62	89.38	71.88	90.67	0.876
	Mean	88.33	92.71	98.33	97.71	93.54	92.50	90.77	91.04	94.58	90.62	82.50	91.87	89.38	76.71	90.71	0.876
ATCNet	1	85.00	85.62	96.25	98.12	88.12	92.50	92.45	86.25	91.25	89.38	75.62	88.75	90.62	78.12	88.43	0.846
	2	82.50	86.25	95.00	96.25	86.25	93.75	88.05	92.50	95.00	86.25	75.00	91.25	93.12	76.88	88.43	0.846
	3	83.75	87.50	96.88	97.50	88.75	94.38	90.57	88.12	89.38	88.75	78.75	89.38	85.00	68.12	87.63	0.835
	Mean	83.75	86.46	96.04	97.29	87.71	93.54	90.36	88.96	91.21	88.13	76.46	89.79	89.58	74.37	88.17	0.842

Note

Multi-seed reporting is particularly relevant in EEG decoding, where outcomes can vary due to stochastic training (initialization, mini-batch ordering, etc.). These tables are meant as a transparency/reproducibility check, aligned with the experimental protocol discussed in the paper.

Reproducibility note. All RatioWaveNet results are obtained using the implementation released in this repository. ATCNet results are recomputed using the official ATCNet repository, without modifying the original architecture. For both models, we adopt the same data split, preprocessing pipeline, and evaluation protocol. Each run differs only by the random seed (affecting weight initialization and training stochasticity), ensuring a fair and reproducible comparison.

---

None vs STFT vs RDWT in RatioWaveNet

Dataset	Preprocessing	Accuracy (%)	Kappa (κ)
BCI Competition IV-2a	None	79.36	72.50
	STFT	79.09	72.10
	RDWT	82.14	76.20
BCI Competition IV-2b	None	97.00	69.80
	STFT	97.23	67.60
	RDWT	97.69	80.60
High-Gamma Dataset	None	87.45	83.30
	STFT	88.57	84.80
	RDWT	92.81	90.40

---

About RatioWaveNet

---

Dataset

This project uses three publicly available EEG motor imagery datasets for training and evaluation:

1. BCI Competition IV – Dataset 2a

• Description: EEG data from 9 subjects performing four different motor imagery tasks: left hand, right hand, feet, and tongue movements. Each subject completed two sessions (training and evaluation), with 288 trials per session.
• Format: .mat files
• Download: BCI Competition IV – Dataset 2a

2. BCI Competition IV – Dataset 2b

• Description: EEG recordings from 9 subjects performing left and right hand motor imagery tasks. The dataset contains five sessions per subject, with three sessions including feedback.
• Format: .gdf files
• Download: BCI Competition IV – Dataset 2b

3. High-Gamma Dataset (HGD)

• Description: EEG recordings from 14 subjects performing motor execution tasks, recorded using 128 channels. This dataset is well-suited for high-frequency EEG analysis.
• Format: .mat files
• Download: Available through the GIN Repository

Note: Each dataset has a dedicated notebook in this repository, which includes download links and preprocessing instructions.