An Intelligent System for Early Detection of Schizophrenia Using EEG-Based Deep Learning Approach

Abstract

SZ (Schizophrenia) is a multifaceted neurological disorder that demands a timely and precise diagnosis to facilitate optimal treatment outcomes. This research introduces a DL-driven framework tailored for classifying SZ through EEG signal analysis, underpinned by advanced preprocessing and feature extraction strategies. To enhance EEG signal clarity,ICA is first applied to eliminate artifacts. Subsequently, CWT is used to extract vital features across both temporal and frequency domains, ensuring that essential data is retained for classification. Three state-of-the-art DL models—TCN, 1DCNN, and LSTM—are implemented to perform classification. Of these, the TCN model, when combined with ICA and CWT, delivers the most robust results. The framework is evaluated using data from the Kaggle MLSP 2014 Schizophrenia Classification Challenge, where the TCN consistently outperforms 1DCNN and LSTM in metrics such as accuracy, precision, and recall. This underscores TCN\'s ability to capture temporal dynamics in EEG signals effectively.Overall, the proposed model stands out as a compelling and efficient solution for automatic schizophrenia detection, providing a promising decision-support tool for clinical applications by merging efficient preprocessing, informative feature extraction, and powerful DL methodologies.

Introduction

???? Background

• Schizophrenia (SZ) is a severe mental disorder affecting ~24 million people worldwide.

• It causes disruptions in cognition, emotion, perception, and behavior, reducing life expectancy by 10–20 years.

• Traditional SZ diagnosis relies on subjective assessments, leading to potential delays or inaccuracies.

???? Need for Advanced Diagnosis

• EEG (Electroencephalography) is a non-invasive, high-temporal-resolution tool suitable for identifying neural abnormalities in SZ.

• However, EEG signals are high-dimensional, non-stationary, and noisy, requiring robust preprocessing and feature extraction.

• Machine Learning (ML) and Deep Learning (DL) approaches have shown promise for automatic SZ classification.

?? Proposed System Overview

Goal:

Develop a DL-based framework for automated SZ classification using preprocessed EEG data.

Dataset:

• Kaggle MLSP 2014 SZ Classification Challenge dataset with EEG recordings from SZ patients and healthy controls.

Workflow:

1. Preprocessing:

   • Use Independent Component Analysis (ICA) to remove artifacts (eye blinks, muscle noise).

   • ICA separates EEG into statistically independent components, improving signal quality.

2. Feature Extraction:

   • Apply Continuous Wavelet Transform (CWT) to capture both time and frequency characteristics of EEG.

3. Classification:

   • Compare performance of three DL models:

     • TCN (Temporal Convolutional Network)

     • 1D-CNN (One-Dimensional Convolutional Neural Network)

     • LSTM (Long Short-Term Memory Network)

   • Models trained to classify EEG as SZ or non-SZ.

???? Results

• TCN model (with ICA + CWT) showed best performance in terms of accuracy, precision, and recall.

  • TCN excels in modeling long-term temporal dependencies and efficient training.

• Validates potential for creating a reliable clinical decision-support system for SZ diagnosis.

???? Related Research Highlights

1. Nivashini Nattudurai (2023)

• Used Butterworth filter + Power Spectral Density (PSD) features.

• LDA classifier achieved 70% accuracy for SZ classification.

2. Guibing Li et al. (2023)

• Proposed a self-attention neural network for MRI-based SZ detection.

• Achieved 76.76% accuracy, identified affected brain regions (thalamus, cerebellum).

3. Zhifen Guo et al. (2021)

• Used CNN on EEG data, achieving 92% accuracy.

• Emphasized deep learning as a viable tool for objective SZ diagnosis.

4. Simone Poetto et al. (2024)

• Used Topological Data Analysis (TDA) for EEG classification.

• Achieved 80% accuracy with only 2 features, showing potential for interpretable and efficient SZ diagnosis.

5. Nadezhda Shanarova et al. (2024)

• Extracted features from event-related potentials (ERPs) in EEG using blind source separation.

• SVM achieved 96.7% sensitivity and 97.7% specificity.

• SHAP analysis validated physiological relevance of features.

6. Enkhmaa Luvsannyam et al. (2022)

• Explored genetic and neurochemical underpinnings of SZ.

• Genes like DTNBP1 and NRG1 linked to SZ susceptibility.

• Treatment primarily pharmacological but challenged by side effects and variability.

????? Technical Components of Proposed System

???? Preprocessing - ICA

• Removes artifacts (e.g., muscle noise) from EEG by estimating independent components.

• Equations:

  • Mixing model: X = A.S

  • ICA goal: estimate S = W.X where W ≈ A?¹

???? Feature Extraction - CWT

• Converts EEG into time-frequency domain using wavelet transformations.

• Equation:
  CWT(a, b) = ∫ x(t)·ψ*((t-b)/a) dt

???? Classifiers Used:

• LSTM:

  • Specialized RNN for modeling long-term dependencies in EEG.

  • Uses gates: forget, input, and output.

• 1D-CNN:

  • Extracts localized temporal features using 1D filters.

  • Convolution equation:
    y[i] = ∑ x[i+j]·w[j] + b

• TCN:

  • Uses dilated causal convolutions to capture long-range dependencies.

  • Fast, parallelizable, and order-preserving.

  • Equation:
    y[t] = ∑ x[t - d·i]·w[i]

Conclusion

This study introduced a DL-driven framework for classifying SZ using EEG signals, integrating multiple stages for improved diagnostic accuracy. The methodology incorporated ICA to eliminate artifacts, followed by CWT to extract rich time-frequency features. These processed features were then input into three classification models: TCN, 1D-CNN, and LSTM networks.Among these, the TCN model emerged as the top performer, achieving an accuracy of 92.20%, precision of 90.50%, and recall of 92.90%, highlighting its ability to effectively capture long-range temporal dependencies in EEG data.The findings underscore that combining advanced preprocessing techniques with strong temporal modeling can significantly boost the performance of automated SZ detection systems. This approach presents a promising solution for early diagnosis and monitoring, offering potential value in clinical decision-support environments.For future research, the framework could be further enhanced by incorporating multi-channel spatial feature fusion, attention mechanisms, or transformer-based architectures to improve both model interpretability and predictive accuracy. Additionally, evaluating the system on larger and more heterogeneous clinical datasets would help improve its generalizability. Another important direction is the real-time integration of this framework with EEG acquisition systems to enable practical deployment in clinical settings.

References

[1] Nivashini Nattudurai (2023), “Classifying Schizophrenia Disorder through EEG Signal Analysis and Machine Learning,\" 2023 IEEE MIT Undergraduate Research Technology Conference (URTC), Cambridge, MA, USA, 2023, pp. 1-4, doi: 10.1109/URTC60662.2023.10535028. [2] Guibing Li; WeidongJin; Rude Lin; Ying Tan&Jiaqing Miao (2023), “A Classification and Pathology Study on Schizophrenia Based on Self-Attention Model,\" 2023 3rd International Symposium on Computer Technology and Information Science (ISCTIS), Chengdu, China, 2023, pp. 562-571, doi: 10.1109/ISCTIS58954.2023.10213043. [3] ZhifenGuo; Lezhou Wu; Yun Li; Beilin Li (2021), “Deep neural network classification of EEG data in schizophrenia,\" 2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS), Suzhou, China, 2021, pp. 1322-1327, doi: 10.1109/DDCLS52934.2021.9455509. [4] Simone Poetto; W?odzis?awDuch (2024), Classification of Schizophrenia EEG recording using homological features,\" 2024 International Joint Conference on Neural Networks (IJCNN), Yokohama, Japan, 2024, pp. 1-8, doi: 10.1109/IJCNN60899.2024.10650296. [5] NadezhdaShanarova, Marina Pronina; Mikhail Lipkovich, Valery Ponomarev; Andreas Müller& JuriKropotov (2024), “Classification of Schizophrenia Patients on the Basis of Latent Components of ERPs using a Machine Learning System,\" 2024 Sixth International Conference Neurotechnologies and Neurointerfaces (CNN), Kaliningrad, Russian Federation, 2024, pp. 198-201, doi: 10.1109/CNN63506.2024.10705876. [6] EnkhmaaLuvsannyam 1,2,?, Molly S Jain 3, Maria Kezia Lourdes Pormento 4, Hira Siddiqui 5, Angela Ria A Balagtas 6, Bernard O Emuze 7, Teresa Poprawski (2022), “Neurobiology of Schizophrenia: A Comprehensive Review”, Cureus, 14(4):e23959. doi: 10.7759/cureus.23959 [7] Rinku Supakar, Parthasarathi Satvaya&, Prasun Chakrabarti (2022), “A deep learning based model using RNN-LSTM for the Detection of Schizophrenia from EEG data”, Computers in Biology and Medicine, Vol. 151, Part A, December 2022, 106225 [8] Bethany Gosala , Avnish Ramvinay Singh, Himanshu Tiwari & Manjari Gupta (2025), “GCN-LSTM: A hybrid graph convolutional network model for schizophrenia classification”Biomedical Signal Processing and Control”, Volume 105, July 2025, 107657 [9] Catalan A, Salazar De Pablo G, Aymerich C, Damiani S, Sordi V, Radua J, et al. Neurocognitive functioning in individuals at clinical high risk for psychosis: a systematic review and meta-analysis. JAMA Psychiatry. 2021;78:859–67. [10] Salazar de Pablo G, Radua J, Pereira J, Bonoldi I, Arienti V, Besana F, et al. Probability of transition to psychosis in individuals at clinical high risk: an updated meta-analysis. JAMA Psychiatry. 2021;78:970–8.

Copyright

Copyright © 2025 Jananee J, Dr. F. Emerson Solomon. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.