Source-linked AI summary
Automatic Diagnosis of Schizophrenia in EEG Signals Using CNN-LSTM Models
Afshin Shoeibi, Delaram Sadeghi, Parisa Moridian, Navid Ghassemi, Jonathan Heras, Roohallah Alizadehsani, Ali Khadem, Yinan Kong, Saeid Nahavandi, Yu-Dong Zhang, Juan M. Gorriz
TL;DR
The paper addresses automated schizophrenia diagnosis from EEG signals by comparing conventional machine learning with deep-learning architectures. It segments and normalizes EEG recordings, evaluates multiple classifiers with five-fold cross-validation, and reports 99.25 ± 0.25 accuracy for its best CNN-LSTM model. The study is limited by small, diagnosis-only datasets and by the absence of separate age- and gender-specific classifiers.
Problem
EEG-based SZ diagnosis is challenging, motivating automated methods that can support quicker and more accurate diagnosis.
Method
The study compares conventional classifiers with LSTM, 1D-CNN, and 1D-CNN-LSTM models on normalized, 25-second EEG frames using a Warsaw dataset.
Results
99.25 ± 0.25 accuracy was obtained by the 13-layer 1D-CNN-LSTM with ReLU activation and Z-Score + L2 normalization.
Takeaways & Limitations
The proposed model obtained higher accuracy than a vast majority of prior SZ-diagnosis studies compared in the paper.
Takeaways & Limitations
The available EEG datasets contain limited numbers of cases, and the dataset was not suitable for prognosis or early diagnosis.
Abstract
from arXiv · showhide
Schizophrenia (SZ) is a mental disorder whereby due to the secretion of specific chemicals in the brain, the function of some brain regions is out of balance, leading to the lack of coordination between thoughts, actions, and emotions. This study provides various intelligent deep learning (DL)-based methods for automated SZ diagnosis via electroencephalography (EEG) signals. The obtained results are compared with those of conventional intelligent methods. To implement the proposed methods, the dataset of the Institute of Psychiatry and Neurology in Warsaw, Poland, has been used. First, EEG signals were divided into 25 s time frames and then were normalized by z-score or norm L2. In the classification step, two different approaches were considered for SZ diagnosis via EEG signals. In this step, the classification of EEG signals was first carried out by conventional machine learning methods, e.g., support vector machine, k-nearest neighbors, decision tree, naïve Bayes, random forest, extremely randomized trees, and bagging. Various proposed DL models, namely, long short-term memories (LSTMs), one-dimensional convolutional networks (1D-CNNs), and 1D-CNN-LSTMs, were used in the following. In this step, the DL models were implemented and compared with different activation functions. Among the proposed DL models, the CNN-LSTM architecture has had the best performance. In this architecture, the ReLU activation function with the z-score and L2-combined normalization was used. The proposed CNN-LSTM model has achieved an accuracy percentage of 99.25%, better than the results of most former studies in this field. It is worth mentioning that to perform all simulations, the k-fold cross-validation method with k = 5 has been used.
I. INTRODUCTION
Schizophrenia diagnosis via EEG is clinically important but challenging because EEG recordings are complex and conventional methods require substantial feature-engineering expertise. This study compares conventional machine-learning and deep-learning approaches using normalized EEG frames from a Warsaw dataset.
- Motivation: EEG is a practical, inexpensive functional neuroimaging modality with high temporal resolution, but long-term multichannel recordings complicate SZ diagnosis.Its signals capture brain electrical activity from the head surface and use multiple channels recorded over extended periods.
- Related work: Automated SZ diagnosis has used conventional machine learning and deep learning, with pipelines typically combining preprocessing, feature extraction or selection, and classification.Conventional features include time, frequency, time-frequency, and nonlinear domains.
- Related work: Deep learning addresses the demanding feature-engineering requirements of conventional methods by learning feature representations through deep layers.The paper cites transfer learning with ResNet-18 as an example of deep feature extraction for EEG-based SZ diagnosis.
- Study aim: The study investigates conventional machine-learning and deep-learning methods for SZ diagnosis using the Institute of Psychiatry and Neurology dataset from Warsaw, Poland.The proposed models include SVM, KNN, decision tree, naïve Bayes, random forest, extremely randomized trees, bagging, LSTM, 1D-CNN, and 1D-CNN-LSTM architectures.
- Study aim: EEG signals are normalized with z-score and L2 methods before classification, and the study evaluates nine LSTM, 1D-CNN, and 1D-CNN-LSTM deep-learning methods.The paper also outlines statistical evaluation and comparison with baseline methods.
II. MATERIAL AND METHODS
The methods section examines the dataset, preprocesses EEG signals, and compares conventional machine-learning baselines with deep-learning models for SZ diagnosis.
- II. MATERIAL AND METHODS: The method proceeds from dataset examination to EEG preprocessing and then to conventional machine-learning and deep-learning models for SZ diagnosis.The section presents the proposed dataset, preprocessing method, and model families in that order.
A. Dataset
The dataset contains multichannel EEG recordings from matched normal individuals and patients with SZ, which are segmented into 25-second frames and normalized before classification.
- Dataset: The dataset includes EEG recordings from 14 normal individuals and patients with SZ, using 19 recorded channels sampled at 250 Hz.The recordings were obtained with eyes closed over 15 minutes using the standard 10-20 electrode configuration.
- Preprocessing: Each overlap-free EEG recording is divided into 25-second frames containing 6250 temporal samples per channel.Each resulting frame has 6250×19 dimensions.
- Preprocessing: The EEG frames are normalized using z-score and L2 methods before conventional machine-learning and deep-learning processing.The paper states that normalization improves model accuracy and performance.
- Baseline models: The conventional baselines include SVM, KNN, decision tree, naïve Bayes, random forest, extremely randomized trees, and bagging.These algorithms are introduced as comparison methods.
3) Decision Tress:
The section describes decision-tree-based classifiers and related ensemble methods, alongside naïve Bayes, using their feature-based classification mechanisms.
- 3) Decision Tress:: A decision tree predicts instance classes by learning simple decision rules from data features.Its binary-tree representation uses input variables and split points, with leaf nodes containing the output variable.
- 4) Na¨ıve Bayes:: Naïve Bayes applies Bayes’ theorem while assuming conditional independence between every pair of features given the class.Class probabilities are computed from feature-specific conditional probabilities and the class prior.
- 5) Random Forest:: Random forest extends bagging by fitting decision trees on dataset subsamples and selecting random feature subsets at each split.Averaging is used to improve predictive accuracy and control over-fitting.
- 6) Extremely Randomized Trees:: Extremely randomized trees form an ensemble of decision trees trained on the whole dataset with randomly selected split points and feature subsets.Unlike random forest, ERT selects split points randomly rather than greedily.
7) Bagging:
The section introduces bagging as an ensemble approach and places the proposed EEG diagnosis models within a broader set of CNN, LSTM, and CNN-LSTM architectures.
- Bagging: Bagging fits base classifiers on random dataset subsets and aggregates their predictions by voting or averaging.Each classifier is created from a different bootstrap sample, allowing repeated observations through sampling with replacement.
- Proposed models: The study examines proposed 1D-CNN, LSTM, and 1D-CNN-LSTM models for schizophrenia diagnosis using EEG signals.
- 1D-CNN models: 1D-CNNs treat signal time as a spatial dimension and are described as lower-cost rivals to recurrent architectures for time-series processing.The models use convolutional, pooling, and fully connected layers.
- 1D-CNN models: The first proposed 1D-CNN model contains nine layers, including convolutional, max-pooling, dropout, flatten, and dense layers.Its convolutional layers use 64 filters with 3×3 dimensions, while ReLU, Leaky ReLU, and seLU activations are compared.
- 1D-CNN models: The second and third 1D-CNN models vary convolutional structure, pooling, dropout, flattening, dense layers, and activation functions.The second model uses three convolutional layers and a kernel size of 2; the third uses two convolutional layers and max pooling with kernel size 2.
2) LSTM models:
The paper presents LSTM architectures for EEG-based schizophrenia diagnosis, including a second design that adds an LSTM layer to examine its effect on accuracy.
- LSTM models: RNNs use feedback and internal memory to process sequences by retaining information from previous inputs.The paper identifies simple RNN, LSTM, and GRU networks as important RNN groups.
- First LSTM model: The first proposed LSTM model consists of six layers with one LSTM layer, dropout layers, and dense layers using ReLU and sigmoid activations.
- Second LSTM model: The second proposed LSTM model consists of seven layers and adds an LSTM layer with kernel size 50 to examine its effect on schizophrenia diagnosis accuracy.
3) CNN-LSTM models:
The CNN-LSTM models combine convolutional feature extraction with recurrent processing for EEG-based schizophrenia diagnosis, with a second version expanding the architecture to 13 layers.
- CNN-LSTM models: CNN-RNN models use convolutional layers first to extract local and spatial EEG patterns before passing outputs to RNN layers.The paper states that convolutional layers extract local and spatial patterns better than RNNs experimentally.
- First CNN-LSTM model: The first proposed CNN-LSTM model has 11 layers, including two convolutional layers, dropout, max pooling, flattening, one LSTM layer, and two dense layers.The dense layers use ReLU and sigmoid activation functions.
- Second CNN-LSTM model: The second proposed CNN-LSTM model has 13 layers, with its first ten layers identical to the previous CNN-LSTM model.Its later layers include a 50-neuron ReLU dense layer, dropout with rate=0.25, and a final layer.
III. STATISTICAL METRICS
The study evaluates its algorithms with 5-fold cross-validation and multiple classification metrics derived from the confusion matrix, together with AUC.
- Cross-validation: 5-fold cross-validation uses all dataset points for both training and testing, and the paper describes the resulting evaluations as more reliable.
- Statistical metrics: Performance is evaluated using accuracy, precision, and recall.The metrics are identified as Acc, Prec, and Rec in the paper.
- Statistical metrics: The area under the ROC curve is also calculated for each method.
IV. EXPERIMENT RESULTS
The experiments compare conventional machine-learning classifiers with CNN, LSTM, and CNN-LSTM deep-learning architectures under different normalization and activation settings. The second CNN-LSTM model with ReLU and combined z-score plus L2 normalization achieved the strongest reported performance.
- Compared methods: The study evaluated KNN, decision tree, SVM, naïve Bayes, bagging, and other conventional classifiers alongside CNN, LSTM, and CNN-LSTM architectures.The deep-learning experiments also varied activation functions, normalization, and regularization settings.
- Conventional machine learning: Bagging with z-score-normalized EEG signals produced the maximum accuracy among the conventional classification algorithms.The conventional methods were also evaluated on raw EEG inputs.
- Activation-function comparisons: The second CNN-LSTM model with Leaky ReLU and combined z-score plus L2 normalization obtained the maximum accuracy in the Leaky ReLU experiments.This result is reported for the proposed deep-learning methods in Table XI.
- Activation-function comparisons: The second LSTM method achieved the maximum accuracy in the experiments using the seLU activation function.These results are presented in Table XII.
- Best-performing configuration: The second CNN-LSTM model with ReLU and combined z-score plus L2 normalization achieved the maximum accuracy across the compared classification methods.ROC curves and learning curves were also reported for the ReLU-based deep-learning models under the normalization settings.
V. LIMITATION OF STUDY
The study’s scope is constrained by limited EEG datasets and by the diagnostic task defined by the available data. The dataset was not suitable for severity assessment, prognosis, or early diagnosis, and classifiers were not separately compared across age and gender groups.
- Dataset scope: Available EEG datasets for schizophrenia diagnosis contain limited numbers of cases, constraining access to EEG- and deep-learning-based diagnostic tools.The authors identify limited case availability as a study limitation.
- Dataset scope: The dataset was used for diagnosis rather than determining schizophrenia severity.The limitation concerns the task supported by the available dataset.
- Dataset scope: The dataset was unsuitable for prognosis or early diagnosis, requiring other appropriate datasets for those purposes.The authors explicitly identify prognosis and early diagnosis as outside this dataset’s suitability.
- Population scope: The classifiers were not separately designed and compared for different age and gender groups.The authors state that other suitable datasets would be needed for this comparison.
- Classification scope: The classifiers were binary, although adding classes for brain disorders with similar symptoms could support multiclass classification.This describes a stated extension of the classifier setup.
VI. CONCLUSION, DISCUSSION, AND FUTURE WORKS
The study compares conventional machine-learning and deep-learning approaches for schizophrenia diagnosis from EEG signals, finding the proposed 1D-CNN-LSTM model achieved the strongest reported accuracy. The authors also identify dataset expansion and broader demographic coverage as future directions.
- The study evaluates automated schizophrenia diagnosis from EEG signals using conventional machine-learning and deep-learning approaches.
- EEG data were divided into 25-second frames and normalized before classification with conventional algorithms and proposed deep-learning models.
- 81.22 ± 1.74 accuracy was obtained by bagging with z-score-normalized EEG, the highest result among the conventional classifiers.
- 99.25 ± 0.25 accuracy was achieved by the 13-layer 1D-CNN-LSTM using ReLU activation and Z-Score + L2 normalization.
- The proposed model achieved higher accuracy than most prior studies and may support quick EEG-based diagnosis as an assistant method in hospitals.
- Future work includes providing EEG datasets for schizophrenia diagnosis and developing models for different age and gender groups.