Identify MRI negative temporal lobe epilepsy with resting fMRI indicators and machine learning techniques.
Authors
Affiliations (3)
Affiliations (3)
- Department of Imaging Center, First Affiliated Hospital of Xinjiang Medical University, Urumqi, 830054, China.
- Department of Imaging Center, First Affiliated Hospital of Xinjiang Medical University, Urumqi, 830054, China. [email protected].
- Department of Medical Plastic and Cosmetic Surgery, First Affiliated Hospital of Xinjiang Medical University, Urumqi, 830054, China. [email protected].
Abstract
About 30% of temporal lobe epilepsy (TLE) cases are negative on MRI, so quantitative diagnosis based on clinical symptoms becomes challenging. There is an urgent need for an accurate and reliable method to differentiate patients with MRI-negative TLE from healthy individuals. This study aimed to explore the use of machine learning methods to diagnose MRI-negative TLE patients based on single and combined resting-state fMRI (rs-fMRI) metrics. This study investigates the diagnostic implications of using both singular and composite resting-state fMRI (rs-fMRI) indices in patients with MRI-negative TLE. We carried out a retrospective analysis of the clinical data and rs-fMRI data of 90 patients with MRI-negative TLE and 90 healthy controls (HCs). Next, the participants were divided into a training set and a test set at 8:2. Functional indices extracted from each brain region included degree centrality (DC), voxel-mirrored homotopic connectivity (VMHC), regional homogeneity (ReHo), fractional amplitude of low-frequency fluctuations (fALFF), and amplitude of low-frequency fluctuations (ALFF). A two-sample t-test was utilized to select significant voxels. After this, classification models based on individual rs-fMRI indices and combined rs-fMRI indices were constructed using ML algorithms such as support vector machines (SVM), random forests (RF), and logistic regression (LR) on the training set. Model performance was evaluated using metrics such as specificity, the area under the receiver operating characteristic curve (AUC), sensitivity, and accuracy, and validations were performed on the test set. Lastly, the feature contribution was assessed using Shapley Additive explanations (SHAP) values. The SVM model employing a combination of rs-fMRI functional indices had optimal performance. On the test set, this model achieved an AUC of 0.89, with an accuracy rate of 82%, where the ALFF values from the cerebellum contributed most significantly to the model. In contrast, ML models based on individual rs-fMRI indices demonstrated inferior classification performance, whereas the RF model using the DC index had the lowest accuracy of 47% on the test set. The SVM model combining the fMRI indices has the greatest potential to distinguish between MRI-negative temporal lobe epilepsy patients and healthy individuals, suggesting a complementary role for the classification of resting-state fMRI indices.