top of page

Results

Logistic Regression and SVM

To evaluate our results, we fit the models using the training data and then did a final evaluation with the test data.  Logistic regression did not perform well and had an overall accuracy of 68%. While SVM performed well with an accuracy of 94%.

 

We used receiver operating characteristic curves (ROC Curves) to compare the model performance. ROC Curves plot the sensitivity, or true positive rate, versus the false positive rate. Plots in the upper left of the graph outperform other models. The ROC curves show that SVM outperformed the logistic regression model.

Logistic Regression

SVM

Logistic Regression

Logistic Regression Test Confusion matri

SVM

SVM Test Confusion matrix.emf.png

Logistic Regression

SVM

Test Results

ROC Curve Test SVM and Logistic.png

Logistic Regression

SVM

  • Test results were consistent with training results: SVM outperformed logistic regression in all metrics.

  • Classifcation accuracy (94% vs 62%), F1 score (93.9 vs 65.0%), and AUC (96.3 vs 50.3%).

Testing Results

Logistic Regression

Logistic Regression Training Confusion m

SVM

SVM Training Confusion matrix.png
ROC Curve Training SVM and Logistic.png

Logistic Regression

SVM

  • SVM had much higher classification accuracy (94% vs 68%) and better F1 score (94.4% vs 70.5%).

  • SVM Area Under the Curve or AUC suggest a much better model than logistic regression (97.8% vs 59.3%).

LSTM

The LSTM model was trained using the 5 category response variable, which we then transformed back to a binary response variable for the test data. LSTM provided an overall accuracy of 99% and the ROC curve further shows that the LSTM model provides near perfect classification of the presence of a seizure.

Training Results

Test Results

LSTM Confusion Matrix Test.png
LSTM ROC Test.png

  • For the test data, the model results were transformed to a binary classification response.

  • The classification accuracy was 99%, with an F1 metric of 99.4% and an AUC of 98%.

LSTM Traing Data Loss vs Accuracy.png

  • The LSTM model was trained using all 5 seizure classification levels.

  • LSTM models were trained using epochs of data to avoid overfitting to the training data

  • Overall classification accuracy for the 5 levels was approximately 75%

Conclusion

Using the LSTM model, SeizureSeeker obtained an overall accuracy of 99% for the detection of seizures.

Logistic Regression and SVM performed well, but the long memory property of LSTM is better suited for the time series EEG data.

Logistic Regression SVM LSTM Confusion M

Access to data makes it difficult to develop models for EEG identification:

  • We used pre-processed data from Rochester Institute of Technology.

​

In the future we will focus on the transformation of live EEG data to the structure used in this research and collaborate with Children's National Health System to expand access to clinical datasets.

The results of this study paves the way for complex classification models such as LSTM to detect epileptic seizures in specific population and may be used in combination with other clinical data to predict neurologic outcomes. In addition, have to potential to assist doctors and improve neurological diagnosis and treatment.

bottom of page