| Sumario: | Objective Constructing a classification model based on 3 nonlinear features of scalp electroencephalography (SEEG) for identifying the preictal phase of epilepsy. Methods Total 83 patients with epilepsy who underwent long - term SEEG monitoring at He'nan Provincial People's Hospital from January 2014 to December 2021 were enrolled. Three nonlinear features [phase lag index (PLI), Lempel- Ziv complexity (LZC) and sample entropy (SampEn)] were calculated. Analysis of variance, rank sum test, and χ2 test were used to complete feature selection of disease predictive value. Single and general models were constructed to identify the preictal phase of epilepsy, respectively, using different machine learning (ML) algorithms including support vector machine (SVM), light gradient boosting machine (LightGBM), and k nearest neighbor (KNN) methods based on five -fold cross-validation. Results The LightGBM model constructed with rank sum test feature selection (α + β + δ + θ + whole) performed best. The average performance in the single model was 0.929 for sensitivity, 0.924 for specificity, 0.934 for accuracy, 0.930 for precision, 0.928 for the F1-score, and 0.971 for the area under the curve (AUC). In the general model, the performance was 0.712 for sensitivity, 0.652 for specificity, 0.682 for accuracy, 0.677 for precision, 0.694 for the F1-score, and 0.766 for AUC. Conclusions The LightGBM model constructed using rank sum test was effective in identifying the preictal phase of epilepsy.
|
|---|