Early Detection of Schizophrenia Using EEG Signals Based on CNN and CNN–LSTM Deep Learning Models

Technical Summary: Early Detection of Schizophrenia Using EEG Signals Based on CNN and CNN–LSTM Deep Learning Models

Problem Statement
Schizophrenia is a severe, chronic psychiatric disorder affecting approximately 24 million people globally, characterized by cognitive, behavioral, and perceptual impairments. Current diagnostic standards rely heavily on subjective clinical interviews and behavioral assessments (e.g., DSM-5, PANSS), which can lead to diagnostic variability, particularly in early stages. While neuroimaging techniques like MRI and PET offer complementary insights, their high cost and limited accessibility hinder routine screening. Electroencephalography (EEG) presents a non-invasive, cost-effective alternative with high temporal resolution. However, automatic schizophrenia detection using EEG remains challenging due to the signals' non-stationary, nonlinear nature, susceptibility to artifacts, and significant inter-subject variability. Conventional machine learning approaches often rely on handcrafted feature extraction, which may fail to capture the complex spatial and temporal dynamics inherent in EEG data.

Methodology
The study proposes and compares two deep learning architectures for automated schizophrenia detection using resting-state EEG signals from a publicly available dataset (81 participants: 49 patients with schizophrenia and 32 healthy controls). The framework consists of five sequential stages:

1. Data Preprocessing: Raw EEG recordings (64 channels, 1024 Hz sampling rate) underwent band-pass filtering (0.5–40 Hz) to remove baseline drift and high-frequency noise. Physiological artifacts (e.g., eye blinks, muscle activity) were removed using Independent Component Analysis (ICA). Signals were downsampled by averaging every eight consecutive samples (reducing trial length from 9216 to 1152 samples) and standardized using Z-score normalization.
2. Model Development:
   • CNN Model: Designed to extract spatial features. It comprises two convolutional layers (64 filters, kernel size 3), ReLU activation, dropout (0.25), max-pooling, and two fully connected layers (100 and 50 neurons) ending with a sigmoid output for binary classification.
   • CNN–LSTM Model: A hybrid architecture designed to capture both spatial and temporal dependencies. It utilizes the same CNN layers for spatial feature extraction, followed by an LSTM layer (100 hidden units) to model long-term temporal relationships in the sequential data. This is followed by two fully connected layers and a sigmoid output.
3. Training and Evaluation: Models were implemented using TensorFlow/Keras. The dataset was split into training, validation, and testing subsets. The Adam optimizer and binary cross-entropy loss function were used with a batch size of 32 over 20 epochs. Early stopping was employed to prevent overfitting. Performance was evaluated using Accuracy, Sensitivity, Specificity, Precision, and Area Under the Receiver Operating Characteristic Curve (AUC).

Key Results
The study demonstrates that integrating temporal sequence learning significantly enhances classification performance:

• CNN Performance: The standalone CNN model achieved a maximum validation accuracy of 73.52%, with a sensitivity of 74.41%, specificity of 70.77%, precision of 88.91%, and AUC of 72.59%. The model exhibited signs of overfitting, as validation accuracy peaked at epoch 17 and subsequently declined while training accuracy continued to rise to 95.43%.
• CNN–LSTM Performance: The hybrid CNN–LSTM model demonstrated superior generalization, achieving a validation accuracy of 92.03%. It attained a sensitivity of 92.49%, specificity of 91.62%, precision of 90.37%, and an AUC of 92.06%. The model showed consistent improvement in both training and validation metrics without the divergence observed in the CNN model.
• Confusion Matrix Analysis: The CNN model struggled to identify schizophrenia cases, correctly classifying only 46.78% of patients, whereas it correctly identified 88.89% of healthy controls. In contrast, the CNN–LSTM model correctly identified 91.32% of schizophrenia patients and 89.47% of healthy controls, indicating a substantial reduction in false negatives and false positives.

Significance and Claims
The authors claim that the proposed CNN–LSTM framework offers a robust, computer-aided diagnostic approach for early schizophrenia screening. The study concludes that while CNNs are effective at extracting spatial EEG features, they are insufficient on their own to capture the complex, dynamic temporal variations associated with schizophrenia. By incorporating LSTM layers, the hybrid model successfully learns long-term temporal dependencies, leading to significantly higher diagnostic accuracy and better generalization.

The paper positions this work as a step toward objective, efficient, and clinically applicable diagnostic systems. However, the authors maintain a modest stance regarding clinical deployment, acknowledging limitations such as the relatively small dataset size (81 participants) and the need for validation on independent, multi-center clinical datasets before practical implementation. They emphasize that the current findings support the potential of deep learning as a supportive tool for psychiatric diagnosis but do not claim the model is ready for standalone clinical use without further rigorous testing.