ShivamKumarPrajapati-123/EEG-Seizure-Detection
GitHub: ShivamKumarPrajapati-123/EEG-Seizure-Detection
基于 CHB-MIT 脑电数据集的端到端癫痫发作检测流水线，综合运用信号处理、特征工程和随机森林分类器，并深入分析了严重类别不平衡对模型性能的影响。
Stars: 0 | Forks: 0
# 基于信号处理和机器学习的脑电癫痫发作检测
## 作者
**Shivam Prajapati**
计算机科学工程 — 人工智能与机器学习
## 1. 项目概述
癫痫发作是一种神经系统事件，会在脑电图 (EEG) 信号中产生异常模式。从 EEG 记录中自动检测癫痫发作是生物医学信号处理和机器学习的一个重要应用。
本项目展示了一个端到端的机器学习 pipeline，利用 **CHB-MIT Scalp EEG Database** 从 EEG 信号中检测癫痫发作活动。
该系统处理原始 EEG 记录，执行信号预处理，将连续的 EEG 信号分割成固定长度的窗口，分配癫痫发作和非发作标签，提取统计和频域特征，并应用 Random Forest 分类器进行二分类。
该项目还评估了严重类别不平衡的影响，并研究了分类阈值调整，以了解癫痫发作敏感度与误报预测之间的权衡。
完整的 pipeline 包括：
- 从 EDF 文件加载 EEG 数据
- EEG 信号预处理
- 固定长度的 EEG 窗口化
- 癫痫发作和非发作标记
- 统计特征提取
- 频域特征提取
- 机器学习数据集构建
- Random Forest 模型训练
- 模型评估
- 分类阈值分析
- EEG 窗口级预测
- 自动生成评估结果
- 结果可视化与报告
本项目作为研究和学习原型开发，旨在获得以下方面的实践经验：
- 生物医学信号处理
- EEG 数据分析
- 特征工程
- 机器学习
- 不平衡分类
- 模型评估
- 可复现的机器学习工作流
## 2. 目标
本项目的主要目标是：
1. 从 EDF 文件加载 EEG 记录。
2. 了解 EEG 信号属性和元数据。
3. 使用频率滤波预处理 EEG 信号。
4. 将连续的 EEG 记录分割成固定长度的窗口。
5. 识别癫痫发作和非发作的 EEG 窗口。
6. 从 EEG 信号中提取统计特征。
7. 使用功率谱密度 (PSD) 提取频域特征。
8. 构建机器学习特征数据集。
9. 训练 Random Forest 分类器。
10. 使用适当的分类指标评估模型性能。
11. 分析严重类别不平衡的影响。
12. 研究分类阈值调整。
13. 对单个 EEG 窗口执行预测。
14. 生成可复现的评估结果和可视化。
## 3. 数据集
本项目使用 **CHB-MIT Scalp EEG Database**，这是一个公开可用的 EEG 数据集，包含来自患有难治性癫痫的几科受试者的 EEG 记录。
EEG 记录以 **EDF (European Data Format)** 文件提供，并使用 **MNE-Python** 库进行处理。
当前项目专注于来自 CHB-MIT 数据集的 EEG 记录，包括对以下内容的分析：
```
chb01_03.edf
You're right. The problem is that your actual README currently contains only Sections 1–3, and you want the entire remaining content starting from Section 3 onward, not another copy of the same beginning.
Your current README ends at:
chb01_03.edf
Do not delete the content you already have. Keep Sections 1–3 exactly as they are, and paste the following directly after chb01_03.edf.
The raw EDF recordings are not included in this repository because of their large file size. They should be downloaded separately from the CHB-MIT Scalp EEG Database and placed locally inside the data/chb01/ directory.
4. Project Workflow
The complete EEG processing and machine learning workflow is:
Raw EEG EDF Files
│
▼
EEG Data Loading using MNE
│
▼
EEG Signal Understanding
│
▼
Signal Preprocessing
│
▼
4-Second EEG Windowing
│
▼
Seizure / Non-Seizure Labeling
│
▼
Feature Extraction
│
├──────────────────────┐
▼ ▼
Statistical Features Frequency-Domain Features
│ │
└───────────┬──────────┘
▼
Machine Learning Dataset
│
▼
Random Forest Classifier
│
▼
Model Evaluation
│
▼
Classification Threshold
Analysis
│
▼
EEG Window Prediction
│
▼
Results and Visualization
5. EEG Signal Preprocessing
The raw EEG signals are loaded from EDF files using the MNE-Python library.
A band-pass filter is applied to retain relevant EEG frequency components and reduce unwanted low-frequency drift and high-frequency noise.
The filtering range used in the current pipeline is:
Low Cutoff Frequency = 0.5 Hz
High Cutoff Frequency = 40 Hz
The preprocessing step prepares the EEG signals for subsequent windowing and feature extraction.
6. EEG Windowing
The continuous EEG recordings are divided into fixed-length windows.
The current pipeline uses:
Sampling Frequency = 256 Hz
Window Duration = 4 seconds
Samples per Window = 1024
Each EEG window contains:
23 EEG Channels × 1024 Samples
The fixed-length EEG windows are treated as individual samples for feature extraction and machine learning.
This approach converts the continuous EEG recording into smaller segments that can be independently analyzed and classified.
7. Seizure Labeling
Each EEG window is assigned a binary classification label:
0 → Normal / Non-Seizure
1 → Seizure
Seizure annotations associated with the EEG recordings are used to identify windows that overlap with seizure activity.
For the current dataset construction experiment:
Total Windows = 900
Normal Windows = 890
Seizure Windows = 10
The resulting dataset contains a severe class imbalance between normal and seizure samples.
This class imbalance is an important factor when evaluating the performance of the machine learning model.
8. Feature Extraction
Features are extracted from each EEG window using statistical and frequency-domain analysis.
The current feature dataset contains eight features for each EEG window.
8.1 Statistical Features
The following statistical features are extracted from the EEG signals:
Mean
Standard Deviation
Variance
These features are calculated from the EEG channels and aggregated to produce window-level statistical representations.
8.2 Frequency-Domain Features
Power Spectral Density (PSD) is calculated using the Welch method.
The following EEG frequency bands are analyzed:
Frequency Band Frequency Range
Delta 0.5–4 Hz
Theta 4–8 Hz
Alpha 8–13 Hz
Beta 13–30 Hz
Gamma 30–40 Hz
The final feature vector for each EEG window contains:
Mean
Std
Variance
Delta
Theta
Alpha
Beta
Gamma
The resulting dataset has the following dimensions:
Features Shape = (900, 8)
Labels Shape = (900,)
The processed feature data is stored in the data/ directory.
9. Machine Learning Model
A Random Forest Classifier is used as the baseline machine learning model.
Random Forest was selected because it:
Works effectively with tabular feature data.
Can model nonlinear relationships.
Does not require extensive feature scaling.
Provides feature importance information.
Is relatively straightforward to train and interpret.
The trained Random Forest model is saved as:
models/random_forest_model.pkl
The model is later loaded for evaluation and individual EEG window prediction.
The current Random Forest implementation is considered a baseline research model and is not a clinically validated seizure detection system.
10. Model Evaluation
The dataset is divided into training and testing subsets using an 80/20 split.
The resulting dataset contains:
Total Samples = 900
Training Samples = 720
Testing Samples = 180
The class distribution in the test set is:
Normal Samples = 178
Seizure Samples = 2
The model is evaluated using the following metrics:
Accuracy
Precision
Sensitivity (Recall)
Specificity
F1-Score
Balanced Accuracy
Confusion Matrix
Classification Report
Because the dataset is severely imbalanced, accuracy alone is not considered sufficient for evaluating seizure detection performance.
11. Baseline Model Results
The Random Forest model is first evaluated using the default classification threshold:
Threshold = 0.50
The resulting confusion matrix is:
Predicted Normal Predicted Seizure
Actual Normal 178 0
Actual Seizure 2 0
The final baseline metrics are:
Metric Result
Accuracy 98.89%
Precision 0.00%
Sensitivity 0.00%
Specificity 100.00%
F1-Score 0.00%
Balanced Accuracy 50.00%
The model predicted all 180 test samples as the majority Normal class.
Although the overall accuracy is high, the model failed to detect any seizure samples.
This demonstrates an important limitation of accuracy in highly imbalanced medical classification problems.
The baseline model therefore cannot be considered effective for seizure detection despite its high overall accuracy.
The baseline confusion matrix is available in:
results/confusion_matrix.png
12. Classification Threshold Analysis
The predicted seizure probabilities were further analyzed by changing the classification threshold.
The following thresholds were evaluated:
0.50
0.20
0.10
0.05
0.01
The results were:
Threshold Sensitivity Specificity
0.50 0.00% 100.00%
0.20 0.00% 100.00%
0.10 0.00% 96.07%
0.05 50.00% 87.64%
0.01 50.00% 63.48%
At a classification threshold of 0.05, the confusion matrix was:
Predicted Normal Predicted Seizure
Actual Normal 156 22
Actual Seizure 1 1
At this threshold:
Total Seizure Samples = 2
Detected Seizures = 1
Sensitivity = 50.00%
However:
False Positives = 22
Specificity = 87.64%
This demonstrates the trade-off between increasing seizure sensitivity and increasing false-positive predictions.
The threshold analysis results are saved in:
results/threshold_analysis.png
The threshold analysis is exploratory because the test set contains only two seizure samples. Therefore, these results should not be considered statistically reliable threshold optimization.
13. Key Findings
The main findings of the current experiment are:
The baseline Random Forest model achieved 98.89% accuracy at the default threshold.
The high accuracy was strongly influenced by severe class imbalance.
The baseline model failed to detect seizure samples at the default threshold.
Baseline sensitivity was 0%.
Baseline balanced accuracy was 50%.
Lowering the classification threshold to 0.05 increased sensitivity to 50% on the current test split.
The improved sensitivity resulted in additional false-positive predictions.
Specificity decreased from 100% to 87.64% at the 0.05 threshold.
The model demonstrates the difficulty of seizure detection when only a small number of seizure examples are available.
Accuracy alone is not an appropriate measure of success for this highly imbalanced classification problem.
The current model is therefore best considered a baseline research prototype demonstrating an end-to-end EEG signal processing and machine learning workflow.
14. Prediction Analysis
The trained Random Forest model was also tested on individual EEG windows.
A normal EEG window was classified as:
Normal / Non-Seizure
Known seizure windows from the analyzed recording were also evaluated individually.
For the known seizure windows:
Window Range = 749–758
Total Seizure Windows = 10
Detected as Seizure = 0
Detected as Normal = 10
This indicates that the current Random Forest model failed to detect the known seizure windows in the analyzed recording.
This result further demonstrates the limitations of the current baseline model and highlights the impact of:
Severe class imbalance
Limited seizure samples
Limited feature representation
Limited training data
Potential differences between EEG recordings
The prediction pipeline is therefore useful for demonstrating the current model's behavior and identifying areas requiring further research.
15. Feature Importance
The Random Forest model provides feature importance estimates that indicate the relative contribution of the extracted features to the model's predictions.
The current feature set contains:
Mean
Std
Variance
Delta
Theta
Alpha
Beta
Gamma
The feature importance visualization is saved in:
results/feature_importance.png
The feature importance analysis provides an initial interpretation of which signal characteristics contribute to the baseline model.
However, feature importance should not be interpreted as evidence that any particular EEG frequency band is clinically diagnostic of seizures.
16. Limitations
16.1 Severe Class Imbalance
The current dataset contains substantially more normal EEG windows than seizure windows.
Normal Windows = 890
Seizure Windows = 10
This imbalance makes it difficult for the model to learn robust seizure-specific patterns.
16.2 Limited Seizure Samples
Only a small number of seizure windows are available in the current experiment.
The test set contains only two seizure samples, making sensitivity estimates highly unstable.
A single additional correct or incorrect seizure prediction can significantly change the measured sensitivity.
16.3 Limited Dataset Scope
The current experiment uses a limited subset of the CHB-MIT dataset.
The model has not yet been extensively evaluated across multiple patients.
16.4 Patient-Independent Generalization
The current experiment does not provide sufficient evidence that the model will generalize to completely unseen patients.
Patient-independent evaluation is required before making stronger claims about generalization.
16.5 Threshold Analysis Limitations
The threshold analysis is based on a very small test set containing only two seizure samples.
Therefore, the threshold results are exploratory and should not be considered statistically reliable threshold optimization.
16.6 No Clinical Validation
This project is an educational and research prototype.
The model has not been clinically validated and should not be used for medical diagnosis, treatment, or clinical decision-making.
17. Future Work
Future improvements may include:
Using a larger number of EEG recordings.
Including data from multiple patients.
Performing patient-independent evaluation.
Increasing the number of seizure samples.
Applying appropriate class imbalance handling techniques.
Exploring class weighting and resampling methods.
Performing stratified cross-validation.
Optimizing classification thresholds using a dedicated validation set.
Evaluating Precision-Recall curves.
Evaluating ROC-AUC.
Testing Support Vector Machines (SVM).
Testing XGBoost and other ensemble methods.
Exploring 1D CNN-based EEG classification.
Exploring LSTM and other deep learning approaches.
Extracting additional time-domain features.
Extracting additional frequency-domain features.
Investigating time-frequency representations such as wavelets.
Evaluating model performance using larger and more balanced test sets.
Improving seizure detection sensitivity while controlling false-positive rates.
Evaluating patient-independent and cross-patient performance.
Investigating subject-specific and generalized seizure detection models.
18. Technologies Used
The project uses the following technologies and libraries:
Python
MNE-Python
NumPy
Pandas
SciPy
Matplotlib
Scikit-learn
Joblib
Jupyter Notebook
19. Project Structure
The repository is organized into separate directories for data, notebooks, models, results, visualizations, and source code.
EEG-Seizure-Detection/
│
├── data/
│ ├── features.csv
│ ├── features.npy
│ ├── labels.npy
│ └── README.md
│
├── images/
│ ├── random_forest_baseline_confusion_matrix.png
│ ├── random_forest_confusion_matrix.png
│ └── random_forest_threshold_confusion_matrix.png
│
├── models/
│ └── random_forest_model.pkl
│
├── notebooks/
│ ├── 01_Reading_EEG.ipynb
│ ├── 02_Understanding_EEG.ipynb
│ ├── 03_Preprocessing.ipynb
│ ├── 04_Windowing.ipynb
│ ├── 05_Labeling.ipynb
│ ├── 06_Feature_Extraction.ipynb
│ ├── 07_Model_Training.ipynb
│ ├── 08_Dataset_Builder.ipynb
│ ├── 09_Model_Evaluation.ipynb
│ └── 10_Prediction.ipynb
│
├── results/
│ ├── confusion_matrix.png
│ ├── feature_importance.png
│ ├── final_evaluation.txt
│ └── threshold_analysis.png
│
├── src/
│ └── final_evaluation.py
│
├── .gitignore
├── README.md
└── requirements.txt
The raw CHB-MIT EDF recordings and the generated windows.npy file are intentionally excluded from the GitHub repository because of their large size.
They should remain available locally for reproducing the full data-processing pipeline.
20. Installation
Clone the repository:
git clone https://github.com/ShivamKumarPrajapati-123/EEG-Seizure-Detection.git
Navigate to the project directory:
cd EEG-Seizure-Detection
Install the required Python packages:
pip install -r requirements.txt
21. Running the Project
The notebooks are designed to be executed in the following general sequence:
01_Reading_EEG
↓
02_Understanding_EEG
↓
03_Preprocessing
↓
04_Windowing
↓
05_Labeling
↓
06_Feature_Extraction
↓
08_Dataset_Builder
↓
07_Model_Training
↓
09_Model_Evaluation
↓
10_Prediction
The notebooks demonstrate the individual stages of the project pipeline.
The final automated evaluation script can be executed from the project root using:
python src/final_evaluation.py
The script generates evaluation outputs and visualizations in the results/ directory.
22. Reproducibility
The project uses fixed random seeds where applicable to improve reproducibility of the machine learning experiments.
The main processed artifacts include:
data/features.csv
data/features.npy
data/labels.npy
models/random_forest_model.pkl
The evaluation outputs include:
results/confusion_matrix.png
results/feature_importance.png
results/final_evaluation.txt
results/threshold_analysis.png
Raw EEG recordings are not stored in the repository and must be obtained separately from the CHB-MIT Scalp EEG Database.
23. Research Status
Status: Research and Development / Baseline Prototype
The current project demonstrates an end-to-end pipeline for EEG signal processing and machine learning-based seizure classification.
The implemented pipeline covers:
EEG data loading
Signal preprocessing
Window segmentation
Seizure labeling
Feature extraction
Dataset construction
Machine learning model training
Model evaluation
Threshold analysis
Individual EEG window prediction
The current experimental results highlight the challenges of seizure detection under severe class imbalance.
The baseline model demonstrates high overall accuracy but poor seizure detection sensitivity, emphasizing the importance of appropriate evaluation metrics and robust dataset construction.
Future development will focus on improving seizure detection sensitivity, increasing the amount of seizure data, performing patient-independent evaluation, and investigating more robust machine learning and deep learning approaches.
24. Disclaimer
This project is developed for educational and research purposes only.
The current model is not a clinically validated medical device and should not be used for medical diagnosis, treatment, or clinical decision-making.
The reported results are based on a limited experimental dataset and should not be interpreted as evidence of clinical effectiveness.
25. Author
Shivam Prajapati
Computer Science Engineering — Artificial Intelligence & Machine Learning
GitHub:
https://github.com/ShivamKumarPrajapati-123You're right. The problem is that your actual README currently contains only Sections 1–3, and you want the entire remaining content starting from Section 3 onward, not another copy of the same beginning.
Your current README ends at:
chb01_03.edf
Do not delete the content you already have. Keep Sections 1–3 exactly as they are, and paste the following directly after chb01_03.edf.
The raw EDF recordings are not included in this repository because of their large file size. They should be downloaded separately from the CHB-MIT Scalp EEG Database and placed locally inside the data/chb01/ directory.
4. Project Workflow
The complete EEG processing and machine learning workflow is:
Raw EEG EDF Files
│
▼
EEG Data Loading using MNE
│
▼
EEG Signal Understanding
│
▼
Signal Preprocessing
│
▼
4-Second EEG Windowing
│
▼
Seizure / Non-Seizure Labeling
│
▼
Feature Extraction
│
├──────────────────────┐
▼ ▼
Statistical Features Frequency-Domain Features
│ │
└───────────┬──────────┘
▼
Machine Learning Dataset
│
▼
Random Forest Classifier
│
▼
Model Evaluation
│
▼
Classification Threshold
Analysis
│
▼
EEG Window Prediction
│
▼
Results and Visualization
5. EEG Signal Preprocessing
The raw EEG signals are loaded from EDF files using the MNE-Python library.
A band-pass filter is applied to retain relevant EEG frequency components and reduce unwanted low-frequency drift and high-frequency noise.
The filtering range used in the current pipeline is:
Low Cutoff Frequency = 0.5 Hz
High Cutoff Frequency = 40 Hz
The preprocessing step prepares the EEG signals for subsequent windowing and feature extraction.
6. EEG Windowing
The continuous EEG recordings are divided into fixed-length windows.
The current pipeline uses:
Sampling Frequency = 256 Hz
Window Duration = 4 seconds
Samples per Window = 1024
Each EEG window contains:
23 EEG Channels × 1024 Samples
The fixed-length EEG windows are treated as individual samples for feature extraction and machine learning.
This approach converts the continuous EEG recording into smaller segments that can be independently analyzed and classified.
7. Seizure Labeling
Each EEG window is assigned a binary classification label:
0 → Normal / Non-Seizure
1 → Seizure
Seizure annotations associated with the EEG recordings are used to identify windows that overlap with seizure activity.
For the current dataset construction experiment:
Total Windows = 900
Normal Windows = 890
Seizure Windows = 10
The resulting dataset contains a severe class imbalance between normal and seizure samples.
This class imbalance is an important factor when evaluating the performance of the machine learning model.
8. Feature Extraction
Features are extracted from each EEG window using statistical and frequency-domain analysis.
The current feature dataset contains eight features for each EEG window.
8.1 Statistical Features
The following statistical features are extracted from the EEG signals:
Mean
Standard Deviation
Variance
These features are calculated from the EEG channels and aggregated to produce window-level statistical representations.
8.2 Frequency-Domain Features
Power Spectral Density (PSD) is calculated using the Welch method.
The following EEG frequency bands are analyzed:
Frequency Band Frequency Range
Delta 0.5–4 Hz
Theta 4–8 Hz
Alpha 8–13 Hz
Beta 13–30 Hz
Gamma 30–40 Hz
The final feature vector for each EEG window contains:
Mean
Std
Variance
Delta
Theta
Alpha
Beta
Gamma
The resulting dataset has the following dimensions:
Features Shape = (900, 8)
Labels Shape = (900,)
The processed feature data is stored in the data/ directory.
9. Machine Learning Model
A Random Forest Classifier is used as the baseline machine learning model.
Random Forest was selected because it:
Works effectively with tabular feature data.
Can model nonlinear relationships.
Does not require extensive feature scaling.
Provides feature importance information.
Is relatively straightforward to train and interpret.
The trained Random Forest model is saved as:
models/random_forest_model.pkl
The model is later loaded for evaluation and individual EEG window prediction.
The current Random Forest implementation is considered a baseline research model and is not a clinically validated seizure detection system.
10. Model Evaluation
The dataset is divided into training and testing subsets using an 80/20 split.
The resulting dataset contains:
Total Samples = 900
Training Samples = 720
Testing Samples = 180
The class distribution in the test set is:
Normal Samples = 178
Seizure Samples = 2
The model is evaluated using the following metrics:
Accuracy
Precision
Sensitivity (Recall)
Specificity
F1-Score
Balanced Accuracy
Confusion Matrix
Classification Report
Because the dataset is severely imbalanced, accuracy alone is not considered sufficient for evaluating seizure detection performance.
11. Baseline Model Results
The Random Forest model is first evaluated using the default classification threshold:
Threshold = 0.50
The resulting confusion matrix is:
Predicted Normal Predicted Seizure
Actual Normal 178 0
Actual Seizure 2 0
The final baseline metrics are:
Metric Result
Accuracy 98.89%
Precision 0.00%
Sensitivity 0.00%
Specificity 100.00%
F1-Score 0.00%
Balanced Accuracy 50.00%
The model predicted all 180 test samples as the majority Normal class.
Although the overall accuracy is high, the model failed to detect any seizure samples.
This demonstrates an important limitation of accuracy in highly imbalanced medical classification problems.
The baseline model therefore cannot be considered effective for seizure detection despite its high overall accuracy.
The baseline confusion matrix is available in:
results/confusion_matrix.png
12. Classification Threshold Analysis
The predicted seizure probabilities were further analyzed by changing the classification threshold.
The following thresholds were evaluated:
0.50
0.20
0.10
0.05
0.01
The results were:
Threshold Sensitivity Specificity
0.50 0.00% 100.00%
0.20 0.00% 100.00%
0.10 0.00% 96.07%
0.05 50.00% 87.64%
0.01 50.00% 63.48%
At a classification threshold of 0.05, the confusion matrix was:
Predicted Normal Predicted Seizure
Actual Normal 156 22
Actual Seizure 1 1
At this threshold:
Total Seizure Samples = 2
Detected Seizures = 1
Sensitivity = 50.00%
However:
False Positives = 22
Specificity = 87.64%
This demonstrates the trade-off between increasing seizure sensitivity and increasing false-positive predictions.
The threshold analysis results are saved in:
results/threshold_analysis.png
The threshold analysis is exploratory because the test set contains only two seizure samples. Therefore, these results should not be considered statistically reliable threshold optimization.
13. Key Findings
The main findings of the current experiment are:
The baseline Random Forest model achieved 98.89% accuracy at the default threshold.
The high accuracy was strongly influenced by severe class imbalance.
The baseline model failed to detect seizure samples at the default threshold.
Baseline sensitivity was 0%.
Baseline balanced accuracy was 50%.
Lowering the classification threshold to 0.05 increased sensitivity to 50% on the current test split.
The improved sensitivity resulted in additional false-positive predictions.
Specificity decreased from 100% to 87.64% at the 0.05 threshold.
The model demonstrates the difficulty of seizure detection when only a small number of seizure examples are available.
Accuracy alone is not an appropriate measure of success for this highly imbalanced classification problem.
The current model is therefore best considered a baseline research prototype demonstrating an end-to-end EEG signal processing and machine learning workflow.
14. Prediction Analysis
The trained Random Forest model was also tested on individual EEG windows.
A normal EEG window was classified as:
Normal / Non-Seizure
Known seizure windows from the analyzed recording were also evaluated individually.
For the known seizure windows:
Window Range = 749–758
Total Seizure Windows = 10
Detected as Seizure = 0
Detected as Normal = 10
This indicates that the current Random Forest model failed to detect the known seizure windows in the analyzed recording.
This result further demonstrates the limitations of the current baseline model and highlights the impact of:
Severe class imbalance
Limited seizure samples
Limited feature representation
Limited training data
Potential differences between EEG recordings
The prediction pipeline is therefore useful for demonstrating the current model's behavior and identifying areas requiring further research.
15. Feature Importance
The Random Forest model provides feature importance estimates that indicate the relative contribution of the extracted features to the model's predictions.
The current feature set contains:
Mean
Std
Variance
Delta
Theta
Alpha
Beta
Gamma
The feature importance visualization is saved in:
results/feature_importance.png
The feature importance analysis provides an initial interpretation of which signal characteristics contribute to the baseline model.
However, feature importance should not be interpreted as evidence that any particular EEG frequency band is clinically diagnostic of seizures.
16. Limitations
16.1 Severe Class Imbalance
The current dataset contains substantially more normal EEG windows than seizure windows.
Normal Windows = 890
Seizure Windows = 10
This imbalance makes it difficult for the model to learn robust seizure-specific patterns.
16.2 Limited Seizure Samples
Only a small number of seizure windows are available in the current experiment.
The test set contains only two seizure samples, making sensitivity estimates highly unstable.
A single additional correct or incorrect seizure prediction can significantly change the measured sensitivity.
16.3 Limited Dataset Scope
The current experiment uses a limited subset of the CHB-MIT dataset.
The model has not yet been extensively evaluated across multiple patients.
16.4 Patient-Independent Generalization
The current experiment does not provide sufficient evidence that the model will generalize to completely unseen patients.
Patient-independent evaluation is required before making stronger claims about generalization.
16.5 Threshold Analysis Limitations
The threshold analysis is based on a very small test set containing only two seizure samples.
Therefore, the threshold results are exploratory and should not be considered statistically reliable threshold optimization.
16.6 No Clinical Validation
This project is an educational and research prototype.
The model has not been clinically validated and should not be used for medical diagnosis, treatment, or clinical decision-making.
17. Future Work
Future improvements may include:
Using a larger number of EEG recordings.
Including data from multiple patients.
Performing patient-independent evaluation.
Increasing the number of seizure samples.
Applying appropriate class imbalance handling techniques.
Exploring class weighting and resampling methods.
Performing stratified cross-validation.
Optimizing classification thresholds using a dedicated validation set.
Evaluating Precision-Recall curves.
Evaluating ROC-AUC.
Testing Support Vector Machines (SVM).
Testing XGBoost and other ensemble methods.
Exploring 1D CNN-based EEG classification.
Exploring LSTM and other deep learning approaches.
Extracting additional time-domain features.
Extracting additional frequency-domain features.
Investigating time-frequency representations such as wavelets.
Evaluating model performance using larger and more balanced test sets.
Improving seizure detection sensitivity while controlling false-positive rates.
Evaluating patient-independent and cross-patient performance.
Investigating subject-specific and generalized seizure detection models.
18. Technologies Used
The project uses the following technologies and libraries:
Python
MNE-Python
NumPy
Pandas
SciPy
Matplotlib
Scikit-learn
Joblib
Jupyter Notebook
19. Project Structure
The repository is organized into separate directories for data, notebooks, models, results, visualizations, and source code.
EEG-Seizure-Detection/
│
├── data/
│ ├── features.csv
│ ├── features.npy
│ ├── labels.npy
│ └── README.md
│
├── images/
│ ├── random_forest_baseline_confusion_matrix.png
│ ├── random_forest_confusion_matrix.png
│ └── random_forest_threshold_confusion_matrix.png
│
├── models/
│ └── random_forest_model.pkl
│
├── notebooks/
│ ├── 01_Reading_EEG.ipynb
│ ├── 02_Understanding_EEG.ipynb
│ ├── 03_Preprocessing.ipynb
│ ├── 04_Windowing.ipynb
│ ├── 05_Labeling.ipynb
│ ├── 06_Feature_Extraction.ipynb
│ ├── 07_Model_Training.ipynb
│ ├── 08_Dataset_Builder.ipynb
│ ├── 09_Model_Evaluation.ipynb
│ └── 10_Prediction.ipynb
│
├── results/
│ ├── confusion_matrix.png
│ ├── feature_importance.png
│ ├── final_evaluation.txt
│ └── threshold_analysis.png
│
├── src/
│ └── final_evaluation.py
│
├── .gitignore
├── README.md
└── requirements.txt
The raw CHB-MIT EDF recordings and the generated windows.npy file are intentionally excluded from the GitHub repository because of their large size.
They should remain available locally for reproducing the full data-processing pipeline.
20. Installation
Clone the repository:
git clone https://github.com/ShivamKumarPrajapati-123/EEG-Seizure-Detection.git
Navigate to the project directory:
cd EEG-Seizure-Detection
Install the required Python packages:
pip install -r requirements.txt
21. Running the Project
The notebooks are designed to be executed in the following general sequence:
01_Reading_EEG
↓
02_Understanding_EEG
↓
03_Preprocessing
↓
04_Windowing
↓
05_Labeling
↓
06_Feature_Extraction
↓
08_Dataset_Builder
↓
07_Model_Training
↓
09_Model_Evaluation
↓
10_Prediction
The notebooks demonstrate the individual stages of the project pipeline.
The final automated evaluation script can be executed from the project root using:
python src/final_evaluation.py
The script generates evaluation outputs and visualizations in the results/ directory.
22. Reproducibility
The project uses fixed random seeds where applicable to improve reproducibility of the machine learning experiments.
The main processed artifacts include:
data/features.csv
data/features.npy
data/labels.npy
models/random_forest_model.pkl
The evaluation outputs include:
results/confusion_matrix.png
results/feature_importance.png
results/final_evaluation.txt
results/threshold_analysis.png
Raw EEG recordings are not stored in the repository and must be obtained separately from the CHB-MIT Scalp EEG Database.
23. Research Status
Status: Research and Development / Baseline Prototype
The current project demonstrates an end-to-end pipeline for EEG signal processing and machine learning-based seizure classification.
The implemented pipeline covers:
EEG data loading
Signal preprocessing
Window segmentation
Seizure labeling
Feature extraction
Dataset construction
Machine learning model training
Model evaluation
Threshold analysis
Individual EEG window prediction
The current experimental results highlight the challenges of seizure detection under severe class imbalance.
The baseline model demonstrates high overall accuracy but poor seizure detection sensitivity, emphasizing the importance of appropriate evaluation metrics and robust dataset construction.
Future development will focus on improving seizure detection sensitivity, increasing the amount of seizure data, performing patient-independent evaluation, and investigating more robust machine learning and deep learning approaches.
24. Disclaimer
This project is developed for educational and research purposes only.
The current model is not a clinically validated medical device and should not be used for medical diagnosis, treatment, or clinical decision-making.
The reported results are based on a limited experimental dataset and should not be interpreted as evidence of clinical effectiveness.
25. Author
Shivam Prajapati
Computer Science Engineering — Artificial Intelligence & Machine Learning
GitHub:
https://github.com/ShivamKumarPrajapati-123
```
标签：Apex, EEG分析, 不均衡分类, 医疗AI, 机器学习, 特征工程, 生物医学信号处理, 癫痫检测, 逆向工具, 随机森林