Method for establishing column diagram model for diagnosis and prediction of different types of temporal lobe epilepsy

Abstract

The invention discloses a method for establishing a column diagram model for diagnosis and prediction of different types of temporal lobe epilepsy. The method comprises the following steps: acquiring clinical data and an MRI image of a temporal lobe epilepsy patient; constructing a clinical feature prediction model by utilizing Logistic regression; preprocessing the MRI image, segmenting a hippocampus region of interest (ROI) and extracting radiomics features, establishing a Radscore model based on screened image features, and combining the Radscore model with clinical features to construct a joint model; performing performance evaluation on the clinical model, the radiomics model and the joint model; and constructing a temporal lobe epilepsy typing prediction column diagram based on the joint model, and realizing prediction result visualization and individualized risk assessment. According to the method, the clinical-image combined model is developed by integrating high-throughput radiomics characteristics and clinical independent factors, high-precision prediction of the temporal lobe epilepsy type is achieved, and the method can provide a reliable basis for clinical early-stage typing diagnosis, precise intervention and personalized treatment.

Description

Technical Field

[0001] This invention belongs to the fields of biomedical engineering and clinical medicine, specifically relating to a method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy. Background Technology

[0002] Epilepsy is the second most common neurological disorder, affecting approximately 70 million people worldwide, with a prevalence of 7.2‰. In my country, the number of patients is close to 10 million. Temporal lobe epilepsy (TLE) accounts for about one-third of focal epilepsy and is the most common type, with seizures originating in the temporal lobe. More than half of TLE patients develop refractory epilepsy (RE). Furthermore, TLE patients often experience cognitive impairment, affecting not only memory but also executive, language, and other cognitive domains, significantly impacting their quality of life. Currently, the diagnosis of comorbid cognitive impairment and refractory epilepsy in TLE patients relies primarily on clinical assessment, lacking reliable early biomarkers and predictive factors.

[0003] Clinically, patients with transient ischemic attack (TLE) typically present with recurrent seizures, and electroencephalography (EEG) shows focal abnormal discharges, but conventional magnetic resonance imaging (MRI) often fails to detect local structural abnormalities. In recent years, with the development of functional imaging techniques, such as functional magnetic resonance imaging (fMRI), positron emission tomography (PET-CT), and single-photon emission computed tomography (SPECT), it is possible to reveal changes in brain microstructure and function to some extent. However, these methods suffer from high detection costs, poor clinical accessibility, and limited resolution, hindering their widespread clinical application. Traditional neuroimaging research focuses primarily on volumetric analysis, diffusion tensor imaging (DTI), or functional imaging, but these largely rely on single parameters and cannot comprehensively reflect the spatial heterogeneity of lesions.

[0004] Radiomics transforms medical images into quantifiable data through high-throughput feature extraction, revealing the spatial complexity of diseases under non-invasive conditions and uncovering potential information that is difficult to identify with the naked eye. In recent years, radiomics has been initially applied in the diagnosis of epilepsy and the localization of epileptogenic foci, showing certain clinical value. For example, existing studies have constructed diagnostic models for hippocampal sclerotic TLE and MRI-negative TLE based on MRI radiomics features, providing a possible approach for clinical auxiliary diagnosis. However, existing studies still have shortcomings: (1) most are retrospective and small-sample studies, with limited reliability and generalization ability of the results; (2) existing simple radiomics models have a single predictive dimension, limiting the model's generalization ability and clinical application value, and the biological significance behind its features is not yet clear; (3) traditional clinical indicators have limited accuracy in classification or predicting disease progression, making it difficult to meet the needs of early and accurate diagnosis. It is evident that existing technologies cannot yet meet the clinical needs for early and accurate classification and prediction of temporal lobe epilepsy. Therefore, there is an urgent need to propose a predictive method for temporal lobe epilepsy classification that combines MRI radiomics with clinical characteristics, and to establish a corresponding model evaluation system, so as to provide new technical means for the early diagnosis and individualized treatment of temporal lobe epilepsy. Summary of the Invention

[0005] The purpose of this invention is to provide a method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy.

[0006] A nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy is established, following these steps:

[0007] (1) Obtain clinical data and MRI images of patients with temporal lobe epilepsy;

[0008] (2) Clinical feature model construction: The collected clinical features were preprocessed and standardized; univariate logistic regression was used to screen and predict potential clinical factors for different types of temporal lobe epilepsy;

[0009] (3) Radiomics feature extraction and Radscore model establishment: The acquired MRI images were preprocessed; the key structures of the temporal lobe were segmented based on the standardized images, and the visualization and manual correction were performed; the shape features, first-order statistical features, texture features and higher-order features were extracted using radiomics software packages, and the features were screened and dimensionality reduced; the Radscore model was constructed based on the screened radiomics features, and logistic regression was used for training and cross-validation.

[0010] (4) Construction of a joint clinical feature and radiomics model: The output of the Radscore model is combined with the selected clinical features; a joint prediction model is constructed and trained and validated by logistic regression to achieve prediction of different types of temporal lobe epilepsy;

[0011] (5) Model performance comparison and evaluation: The performance of clinical models, radiomics models and combined models is evaluated; statistical comparison and validation of models are performed to quantify the performance differences and potential clinical application value between different models;

[0012] (6) Visualization of prediction results and individualized prediction: Based on the joint model, a nomogram for the diagnosis and prediction of different types of temporal lobe epilepsy is constructed; the prediction results are visualized and individualized risk prediction is provided for each patient to assist clinicians in making early subtype diagnosis and treatment decisions.

[0013] The clinical data mentioned in step (1) includes demographic information, disease course, epileptic seizure type and frequency, comorbid history, and cognitive-emotional scale scores.

[0014] The MRI images mentioned in step (1) are the brain MRI images of the patient after admission, including axial 3D-T1 weighted images, T2 weighted images, and T2 fluid attenuation inversion recovery sequences, which are exported in DICOM format for subsequent image processing.

[0015] Step (2) describes the process of predicting potential clinical factors for different types of temporal lobe epilepsy as follows: For each clinical feature X i A univariate logistic regression model was constructed to predict the type Y of temporal lobe epilepsy.

[0016]

[0017] For the screened potential clinical factors X i1 ,X i2 ,...,X ik Perform multivariate logistic regression analysis to establish a multivariate clinical characteristic prediction model:

[0018]

[0019] Where P(Y=1) represents the probability of developing a certain type of temporal lobe epilepsy; β0 is the intercept term; β1,...,β k represents the regression coefficients for each clinical feature.

[0020] The preprocessing described in step (3) includes motion correction, non-brain tissue stripping, intensity normalization, and spatial registration.

[0021] The key temporal lobe structure mentioned in step (3) is the bilateral hippocampus; the radiomics software package is pyRadiomics; the feature selection and dimensionality reduction include K-optimal, recursive feature elimination, and the LASSO method. For the K-optimal method, the k features with the highest correlation to category Y (epilepsy type) are selected: For LASSO dimensionality reduction, key features are selected using L1 regularization: Where θ(·) is the loss function and λ is the regularization coefficient.

[0022] The performance evaluation in step (4) includes calculating the subject operating characteristic curve and area under the curve, sensitivity, specificity, accuracy, and F1 score.

[0023] In the nomogram described in step (6), the first row is the score scale, with a score range of 0 to 100; the second row is the age, with a score range of 15 to 85; the third row is the Radscore, with a score range of -1.6 to 1.4; the fourth row is the total score, with a score range of 0 to 120; the fifth row is the linear prediction value, with a range of -9 to 5; and the sixth row is the prediction probability, with a range of 0.1 to 0.9.

[0024] In the nomogram described in step (6), the first row is the score scale, with a score range of 0 to 100; the second row is the epilepsy type, with a score range of 0 to 1; the third row is the Radscore value, with a score range of -1.4 to 1 point; the fourth row is the total score, with a score range of 0 to 120 points; the fifth row is the linear prediction value, with a range of -7 to 7; and the sixth row is the prediction probability, with a range of 0.1 to 0.9.

[0025] The beneficial effects of this invention are as follows: This invention integrates high-throughput radiomics features with clinically independent factors to develop and validate a combined clinical-imaging model for predicting temporal lobe epilepsy types; it provides a full-process, high-precision, and repeatable method for extracting and screening radiomics features, ensuring model robustness. Results demonstrate that the combined model exhibits superior predictive performance and clinical application value compared to clinical models or single imaging models in multi-level validation, and enables individualized diagnosis through visualized nomograms, providing a reliable basis for early, precise intervention and personalized treatment. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the evaluation method for predictive models for the diagnosis of different types of epilepsy.

[0027] Figure 2 A flowchart of radiomics analysis for evaluating predictive model methods for the diagnosis of different types of epilepsy.

[0028] Figure 3 A schematic diagram comparing the performance of models in the evaluation method of predictive model for the diagnosis of co-occurring cognitive impairment in temporal lobe epilepsy.

[0029] Figure 4 Nodal plot in the evaluation method of predictive model for the diagnosis of temporal lobe epilepsy comorbid with cognitive impairment.

[0030] Figure 5Correction curves in the evaluation method of predictive model for the diagnosis of temporal lobe epilepsy comorbid with cognitive impairment.

[0031] Figure 6 Decision curve analysis diagram in the evaluation method of predictive model for the diagnosis of temporal lobe epilepsy comorbid with cognitive impairment.

[0032] Figure 7 A schematic diagram comparing the performance of models in the evaluation methods for predictive models for the diagnosis of refractory temporal lobe epilepsy.

[0033] Figure 8 Nonograph in the evaluation method of predictive model for the diagnosis of refractory temporal lobe epilepsy.

[0034] Figure 9 Correction curves are used in the evaluation of predictive models for the diagnosis of refractory temporal lobe epilepsy.

[0035] Figure 10 Decision curve analysis diagram in the evaluation method of predictive model for the diagnosis of refractory temporal lobe epilepsy. Detailed Implementation

[0036] To facilitate understanding of the present invention, a more comprehensive description will be given below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0037] Example 1

[0038] Study participants: Patients diagnosed with temporal lobe epilepsy (TLE) at the Department of Neurology, First Affiliated Hospital of Dalian Medical University, between February 2018 and July 2024, aged 18 to 65 years, regardless of gender, were included. They were divided into a TLE-CI group (with cognitive impairment) and a TLE-NCI group (without cognitive impairment) based on the presence or absence of co-occurring cognitive impairment. They were also divided into a TLE-RE group (treatment-resistant epilepsy) and a TLE-DSE group (drug-sensitive epilepsy) group based on the efficacy of antiepileptic drugs (ASMs). This study has been approved by the Ethics Committee of the First Affiliated Hospital of Dalian Medical University, and all patients / legal representatives signed written informed consent forms.

[0039] Inclusion criteria: Patients met the 2017 International League Against Epilepsy (ILAE) classification and diagnostic criteria for epilepsy and epilepsy syndromes; epilepsy was confirmed by complete clinical symptomatology, long-term video electroencephalography (vEEG), and cranial MRI, with the discharge originating in the temporal lobe, with or without secondary generalized tonic-clonic seizures; all enrolled patients were right-handed and able to cooperate in completing cranial MRI, vEEG monitoring, and cognitive-emotional scale examinations; informed consent was obtained from the patient or their family, and the study was approved by the hospital's medical ethics committee. Treatment-resistant epilepsy met the 2010 ILAE diagnostic criteria, i.e., failure to achieve seizure-free status despite the use of two correctly selected and well-tolerated antiepileptic drugs (single or combination therapy); drug-sensitive epilepsy refers to patients whose seizures are effectively controlled after treatment with antiepileptic drugs. All enrolled patients with drug-sensitive and treatment-resistant epilepsy underwent cerebrospinal fluid and blood tests for autoimmune encephalitis and demyelinating antibody series to rule out the possibility of acute symptomatic seizures of autoimmune encephalitis.

[0040] Exclusion criteria: clinical symptom examination or electroencephalogram showing epileptiform discharges outside the temporal lobe; other serious medical or mental illnesses; other central nervous system diseases (such as cerebrovascular diseases, spinal cord lesions, intracranial infections, intracranial space-occupying lesions, etc.); previous traumatic brain injury or surgery; inability to undergo cranial MRI (such as the presence of metal implants); the patient and their family are unable or unwilling to provide written informed consent.

[0041] Research Methods

[0042] The process for evaluating predictive models for the diagnosis of different types of epilepsy is as follows: Figure 1 As shown.

[0043] 1. General Data Collection

[0044] Based on the patients' electronic medical records, we collected all patients' demographic data (including age, sex, education level, etc.), clinical manifestations (disease course, type, frequency and severity of epileptic seizures), comorbidities, multimodal MRI, long-term vEEG, cognitive emotion scale, lumbar puncture cerebrospinal fluid and blood tests, treatment and prognosis, etc.

[0045] 2. Long-term video electroencephalography (vEEG) data acquisition

[0046] This study used a 32-channel Japanese EEG-1200C optical coherent electroencephalography (EEG) system, employing the international 10-20 system for scalp EEG electrode placement. Bilateral auricular electrodes were selected as reference electrodes. Lead positions were FP1, FP2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz, and FPz. The impedance of each lead was <5kΩ, the sampling frequency was 200Hz, the filtering range was 1–70Hz, the sensitivity was 100μV / cm, the time reference was 30mm / s, and the monitoring time was ≥2 hours. Results were jointly interpreted by two experienced EEG physicians.

[0047] 3. Calculation of the interictal discharge index (semi-quantitative) on electroencephalogram (EEG)

[0048] Two experienced EEG physicians interpreted the long-term vEEG to determine the epileptiform discharge index (semi-quantitative) between epileptic seizures, including normal, occasional, few, slightly more, more, and large.

[0049] 4. Cognitive Emotion Scale

[0050] The Montreal Cognitive Assessment (MoCA) is used to assess cognitive function. It includes 11 items across 8 cognitive domains: memory, language, calculation and orientation, attention, visuospatial function, executive function, and abstract thinking. The total score is 30 points, with a score of ≥26 indicating normal cognitive function. If the number of years of education is ≤12 years, the cutoff score for normal cognitive function is 25 points.

[0051] Mini-Mental State Examination (MMSE): A test used to assess cognitive function. A total score of 27-30 indicates normal cognitive function. A score <27 indicates cognitive impairment. Dementia classification criteria (varying according to education level): Illiterate (no education): ≤17 points; Primary school level (≤6 years of education): ≤20 points; Secondary school or above (>6 years of education): ≤24 points; Dementia severity classification: Mild: MMSE ≥21 points; Moderate: MMSE 10-20 points; Severe: MMSE ≤9 points.

[0052] Hamilton Anxiety Rating Scale (HAMA): Used to assess the severity of anxiety symptoms. Total score ≥30 indicates severe anxiety; 22-29 indicates significant anxiety; 15-21 indicates anxiety; 8-14 indicates possible anxiety; ≤7 indicates no anxiety symptoms.

[0053] The Hamilton Depression Rating Scale (HAMD) has 17 items: it is used to assess the severity of depressive symptoms. A total score of ≥24 points may indicate severe depression; 18-23 points may indicate mild to moderate depression; 17-8 points may indicate depressive symptoms; and ≤7 points indicate no depressive symptoms.

[0054] 5. Multimodal MRI data acquisition

[0055] This study used a Philips Ingenia CX 3.0T MRI scanner, a 32-channel phased array coil, a GESigna 1.5T echo-speed MRI scanner, and an orthogonal head coil in the MRI room of the First Affiliated Hospital of Dalian Medical University. Scanning sequences included axial 3D-T1 weighted images (3D-T1WI), T2 weighted images (T2WI), and T2 fluid attenuation inversion recovery sequence (FLAIR). Scanning parameters: The 3D-T1 scan used a brain volume scan sequence (BRAVO) with the following parameters: repetition time (TR) = 8.9 ms, echo time (TE) = 3.5 ms, inversion time (TI) = 450 ms, flip angle (FA) = 13°, field of view (FOV) = 256 mm × 256 mm, matrix = 256 × 256, slice thickness = 1 mm, slice spacing = 0, and approximately 180 slices.

[0056] 6. Radiomics analysis

[0057] The radiomics analysis workflow for evaluating predictive models for the diagnosis of different types of epilepsy is as follows: Figure 2 As shown.

[0058] 6.1 Image Acquisition: 3D-T1WI images were exported from the Picture Archive and Communication System (PACS) in DICOM format for image normalization and ROI (bilateral hippocampus) segmentation.

[0059] 6.2 Image Preprocessing and ROI Segmentation: FreeSurfer 7.2.0 software was used on the Linux platform to perform automated whole-brain processing on 3D-T1WI image data. The recon-all command was used to complete standard preprocessing steps such as motion correction, non-brain tissue and skull dissection, automatic Talairach transformation, gray / white matter segmentation, intensity normalization, and standardized spatial registration. Based on the Desikan-Killiany atlas, subcortical structures were partitioned and labeled, and standardized spatial segmentation results (aseg.mgz, 1mm resolution) were output. 3 The matrix is ​​256×256×256. Subsequently, the mri_extract_label command in FreeSurfer software was used to extract bilateral hippocampal masks (left hippocampus 17, right hippocampus 53) from the segmentation results. The masks were then overlaid onto the normalized spatial image using the FreeView tool for visual inspection and manual correction, ultimately generating a high-precision hippocampal ROI mask for subsequent analysis.

[0060] 6.3 Feature Extraction and Dimensionality Reduction: Radiomic Features: Based on the bilateral hippocampal segmentation results, radiomic features were extracted from each ROI using the Python software pyRadiomics toolkit, including 14 shape features, 18 first-order features, 75 texture features, and 1786 higher-order features, for a total of 1893 features. The enrolled patients were randomly assigned to the training and test sets at an 8:2 ratio, and the data were standardized using Z-score regression analysis. Dimensionality reduction and screening were performed on the standardized data, mainly using the following three methods: (1) K-optimal: Features with strong correlation (correlation coefficient > 0.9) were excluded through Spearman correlation analysis to reduce data redundancy. This method selected the k features with the highest correlation to category Y (epilepsy type): (2) Recursive feature elimination: By repeatedly training the model and gradually removing features that contribute the least to the model, the most feature subset is ultimately retained, improving model performance. (3) Using the Least Absolute Shrinkage and Selection Operator (LASSO), features with the lowest correlation, lowest redundancy, and highest discriminative power are retained for subsequent model construction. LASSO dimensionality reduction selects key features through L1 regularization: Where θ(·) is the loss function and λ is the regularization coefficient.

[0061] Clinical characteristics: The included clinical characteristics were screened for independent clinical factors predicting different types of temporal lobe epilepsy using univariate (P < 0.1) and multivariate logistic regression (P < 0.05) analyses. For each clinical characteristic X... i A univariate logistic regression model was constructed to predict the type Y of temporal lobe epilepsy.

[0062]

[0063] For the screened potential clinical factors X i1 ,X i2 ,…,X ik Perform multivariate logistic regression analysis to establish a multivariate clinical characteristic prediction model:

[0064]

[0065] Where P(Y=1) represents the probability of developing a certain type of temporal lobe epilepsy; β0 is the intercept term; β1,...,β k represents the regression coefficients for each clinical feature.

[0066] 6.4 Construction of clinical feature models, radiomics models, and combined clinical feature-radiomics models:

[0067] 6.4.1 Construction of TLE-CI and TLE-NCI models

[0068] Radscores were obtained by weighting the residual features from radiomics analysis. A radiomics model was constructed using logistic regression, and the model was validated using a 5-fold cross-validation method. A clinical feature model was constructed using logistic regression based on the selected clinical independent predictors. A joint model was constructed using logistic regression based on Radscores and clinical independent predictors.

[0069] 6.4.2 Construction of TLE-RE and TLE-DSE models

[0070] The radscore was obtained by weighting the final residual features from radiomics analysis. A radiomics model was constructed using Support Vector Machines (SVM), and the model was validated using a 5-fold crossover method. A clinical feature model was constructed using SVM based on the selected clinical independent predictors. A joint model was constructed using SVM based on the radscore and clinical independent predictors.

[0071] 6.5 Model Efficacy Evaluation: Based on the machine learning classifier training data, receiver operating characteristic (ROC) curve analysis was used to evaluate the diagnostic efficacy of the models, and the area under the ROC curve (AUC) was calculated for quantitative analysis. The diagnostic efficacy of different models was evaluated using 95% confidence intervals, sensitivity, specificity, accuracy, and F1 score. The Delong test was used to evaluate the differences in diagnostic efficacy between models. Calibration curves were plotted to visually evaluate the predictive accuracy of the models. The Hosmer-Lemersho (HL) test was used to evaluate the goodness of fit of the models. Decision curve analysis (DCA) was applied to evaluate the clinical application value of the models. A nomogram of the combined models was plotted to calculate the probability of predicting temporal lobe epilepsy with cognitive impairment.

[0072] 7. Statistical Analysis

[0073] Statistical analysis was performed using R (3.4.3). The normality of continuous data was assessed using the Shapiro-Wilk test (for data with <50 participants) or the Kolmogorov-Smirnov test (for data with ≥50 participants). Normally distributed data were expressed as mean ± standard deviation, and independent samples t-tests were used for comparisons between two groups. Non-normally distributed data were expressed as median (interquartile range), and Mann-Whitney U test was used for comparisons between two groups. Chi-square tests or Fisher's exact test were used for comparisons of categorical and count data. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was calculated to evaluate the predictive power of each model. The HL test and calibration curves were used to evaluate model goodness of fit, and the Delong test was used to compare differences in ROC diagnostic efficacy. A p-value < 0.05 was defined as statistically significant.

[0074] Research findings:

[0075] 1. Temporal lobe epilepsy with cognitive impairment (TLE-CI) group vs. temporal lobe epilepsy without cognitive impairment (TLE-NCI) ​​group

[0076] Comparison of clinical characteristics between temporal lobe epilepsy with cognitive impairment (TLE-CI) and temporal lobe epilepsy without cognitive impairment (TLE-NCI). This study included 102 TLE patients, divided into the TLE-CI group (75 cases) and the TLE-NCI group (27 cases) based on whether they had comorbid cognitive impairment. First, the demographic and clinical characteristics of the two groups were compared. The results showed that compared to the TLE-NCI group, the TLE-CI group had a longer age of onset (47.05±15.15 years vs. 36.59±12.37 years, P<0.01). There was a statistically significant difference in years of education between the two groups. There were no statistically significant differences between the two groups in terms of gender, disease duration, seizure type, presence of hippocampal sclerosis on MRI, semi-quantitative interictal discharge index on EEG, and treatment efficacy (Table 1).

[0077] Table 1. Comparison of clinical characteristics between temporal lobe epilepsy with cognitive impairment (TLE-CI) and temporal lobe epilepsy without cognitive impairment (TLE-NCI) ​​groups.

[0078]

[0079] Establishment and validation of a predictive model for co-occurring cognitive impairment in temporal lobe epilepsy

[0080] Training set vs. test set: 102 TLE patients were randomly assigned to the training set (81 cases) and the test set (21 cases) in an 8:2 ratio. There were no statistically significant differences in any demographic and clinical characteristics between the training set and the test set (see Table 2).

[0081] Table 2 Comparison of demographic and clinical characteristics between the training and test sets.

[0082]

[0083] Radiomics Model: After data standardization using Z-score regression analysis, K-optimal, recursive feature elimination, and LASSO were employed for data filtering and dimensionality reduction, retaining the 10 features with the lowest correlation, lowest redundancy, and highest discriminative power. A radiomics model was constructed based on these features. In the training set, the AUC of the radiomics model was 0.898 (95% CI: 0.820–0.957), and in the test set, the AUC was 0.767 (95% CI: 0.481–1.000) (see...). Figure 3 (Table 3).

[0084] Clinical Feature Model: Twelve clinical features (gender, age, years of education, disease duration, seizure type, MoCA, MMSE, HAMA, HAMD, treatment efficacy, semi-quantitative long-term vEEG interictal discharge index, and presence or absence of hippocampal sclerosis on head MRI) were analyzed univariately (P < 0.1) and multivariately (P < 0.05) using logistic regression analysis. Age was ultimately identified as an independent predictor of temporal lobe epilepsy with cognitive impairment. A clinical feature model was established based on logistic regression. The AUC value of the clinical feature model was 0.688 (95% CI: 0.569–0.815) on the training set and 0.756 (95% CI: 0.519–0.939) on the test set. (See attached data.) Figure 3 Table 3.

[0085] Joint Model: A joint clinical feature-radiomics model was established based on independent factors of the selected clinical features and Radscore. In the training set, the joint model AUC was 0.903 (95% CI: 0.808–0.969), and in the test set, the joint model AUC was 0.856 (95% CI: 0.467–1.000). See [link to relevant documentation]. Figure 3 Table 3. The nomogram (integrating Radscore and age) visualizes the joint model, as shown below. Figure 4 As shown.

[0086] Performance comparison of the three models: The performance of the three models is compared as shown in Table 3-4. Figure 3 As shown, in the training set, compared with the clinical feature model (AUC: 0.688 (95% CI: 0.569-0.815)), the radiomics model (AUC: 0.898 (95% CI: 0.820-0.957)) had higher efficacy (P = 0.001), and the difference was statistically significant (Delong test, P < 0.05). Compared with the combined model (AUC: 0.903 (95% CI: 0.808-0.969)), the combined model had higher efficacy (P < 0.001), and the difference was statistically significant (Delong test, P < 0.05). In the test set, there were no statistically significant differences between any of the clinical feature model, radiomics model, and combined clinical feature-radiomics model (Delong test, P > 0.05); as shown in the calibration curve (see...). Figure 5 As shown in the figure, the three models mentioned above all demonstrated high value in predicting cognitive impairment comorbid with temporal lobe epilepsy on both the training and test sets. HL test analysis (see...) Figure 5 The results showed that the clinical feature-image combined model had a good fit and high consistency between predicted and actual values ​​(P>0.05); the decision curve (see...) Figure 6 This indicates that each model has high clinical application value. The joint model has high clinical value in most threshold probability ranges.

[0087] Table 3. Comparison of diagnostic efficacy of different models for temporal lobe epilepsy comorbid with cognitive impairment.

[0088]

[0089]

[0090] Abbreviations: AUC: Area under the Receiver Operating Characteristic (ROC) curve; CI: Confidence Interval

[0091] Table 4. Delong test results comparing the effectiveness of each omics model.

[0092] Model training set test set Radiomics Models vs. Clinical Models <![CDATA[0.001 * ]]> 0.959 Radiomics Models vs. Clinical-Imaging Models 0.780 0.293 Clinical-Imaging Model VS Clinical Model <![CDATA[<0.001 * ]]> 0.499

[0093] Note: The data in the table are the results of the De Long test; p-values ​​* p < 0.05

[0094] 2. Treatment-resistant temporal lobe epilepsy (TLE-RE) group vs. drug-sensitive temporal lobe epilepsy (TLE-DSE) group

[0095] Comparison of clinical characteristics between the treatment-resistant temporal lobe epilepsy (TLE-RE) group and the drug-sensitive temporal lobe epilepsy (TLE-DSE) group. 102 patients with temporal lobe epilepsy were divided into the TLE-RE group (38 cases) and the TLE-DSE group (64 cases) based on their ASM efficacy. The demographic and clinical characteristics of the two groups were compared. The results showed that, compared with the TLE-DSE group, the TLE-RE group had a significantly higher proportion of patients with two or more seizure types (P < 0.001). There were no statistically significant differences in gender, age, years of education, disease duration, presence or absence of hippocampal sclerosis, semi-quantitative interictal discharge index on EEG, or cognitive function (see Table 5).

[0096] Table 5. Comparison of demographic and clinical characteristics between the treatment-resistant temporal lobe epilepsy (TLE-RE) group and the drug-sensitive temporal lobe epilepsy (TLE-DSE) group.

[0097]

[0098]

[0099] Establishment and validation of a predictive model for refractory temporal lobe epilepsy

[0100] Training set vs. test set

[0101] One hundred and two TLE patients were randomly assigned to a training set (81 patients) and a test set (21 patients) in an 8:2 ratio. There were no statistically significant differences in any demographic and clinical characteristics between the training and test sets (see Table 6).

[0102] Table 6 Comparison of demographic and clinical characteristics between the training and test sets.

[0103]

[0104] Radiomics Model: After data standardization using Z-score regression analysis, K-optimal, recursive feature elimination, and LASSO were employed for data filtering and dimensionality reduction, retaining 14 features with the lowest correlation, lowest redundancy, and highest discriminative power. A radiomics model was constructed based on these features. In the training set, the AUC of the radiomics model was 0.839 (95% CI: 0.743–0.917); in the test set, the AUC was 0.750 (95% CI: 0.461–0.982). See [link to relevant documentation]. Figure 7 Table 7.

[0105] Clinical Feature Model: Twelve clinical features (gender, age, years of education, disease duration, seizure type, MOCA, MMSE, HAMA, HAMD, treatment efficacy, semi-quantitative long-term vEEG interictal discharge index, and presence or absence of hippocampal sclerosis on head MRI) were analyzed using univariate (P < 0.1) and multivariate logistic regression (P < 0.05). The number of seizure types was ultimately identified as an independent predictor of refractory temporal lobe epilepsy. A clinical feature model was built based on support vector machine. The AUC value of the clinical feature model was 0.628 (95% CI: 0.528–0.726) on the training set and 0.822 (95% CI: 0.635–0.962) on the test set. (See [link to relevant documentation]). Figure 7 Table 7.

[0106] Joint Model: A joint clinical feature-radiomics model was built based on independent factors of selected clinical features and RadScore. In the training set, the joint model AUC was 0.861 (95% CI: 0.771–0.931), and in the test set, the joint model AUC was 0.865 (95% CI: 0.648–1) (see [link to test model]). Figure 7 (Table 7). The nomogram (integrating radscore and number of seizure types) visualizes the joint model, such as... Figure 8 As shown.

[0107] Comparison of model performance: The performance of the three models is compared, as shown in Tables 7-8. Figure 7As shown, in the training set, both the joint model (AUC: 0.861 (95% CI: 0.771–0.931)) and the radiomics model (AUC: 0.839 (95% CI: 0.743–0.917)) showed significantly higher performance than the clinical characteristics model (AUC: 0.628 (95% CI: 0.528–0.726)) (Delong test, P < 0.001). In the test set, the joint model (AUC: 0.865 (95% CI: 0.648–1)) demonstrated significantly higher performance than the radiomics model (AUC: 0.750 (95% CI: 0.461–0.982)) (Delong test, P < 0.05); as shown in the calibration curve (see...). Figure 9 As can be seen, the three models mentioned above all showed high predictive value for refractory temporal lobe epilepsy in both the training and test sets. HL test analysis showed that the clinical feature-imaging combined model had a good fit and high consistency between predicted and actual values ​​(P > 0.05). Decision curves (see...) Figure 10 The results show that each model has high clinical application value, and the combined model has higher clinical value in most threshold probability ranges.

[0108] Table 7. Comparison of diagnostic efficacy of different models for refractory temporal lobe epilepsy.

[0109]

[0110] Abbreviations: AUC: Area under the Receiver Operating Characteristic (ROC) curve; CI: Confidence Interval

[0111] Table 8. Delong test results comparing the efficacy of different omics models.

[0112] Model training set test set Radiomics Models vs. Clinical Models <![CDATA[0.004 * ]]> 0.683 Clinical-Imaging Models vs. Radiomics Models 0.333 <![CDATA[0.035 * ]]> Clinical-Imaging Model VS Clinical Model <![CDATA[<0.001 * ]]> 0.750

[0113] Note: The data in the table are the results of the De Long test; p-values ​​* p < 0.05

[0114] In summary, by extracting radiomic and clinical features from the bilateral hippocampus of patients with temporal lobe epilepsy (TLE), we established clinical feature models, radiomic models, and a combined clinical feature-radiomics model for comorbid cognitive impairment, treatment-resistant temporal lobe epilepsy, and TLE. The combined model demonstrated higher efficacy and greater clinical value. The established clinical-radiomics nomogram serves as a visual and personalized guideline, aiding clinicians in the early and accurate diagnosis of temporal lobe epilepsy.

[0115] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy, characterized in that, Follow these steps: (1) Obtain clinical data and MRI images of patients with temporal lobe epilepsy; (2) Clinical feature model construction: The collected clinical features were preprocessed and standardized; univariate logistic regression was used to screen and predict potential clinical factors for different types of temporal lobe epilepsy; (3) Radiomics feature extraction and Radscore model establishment: The acquired MRI images were preprocessed; the key structures of the temporal lobe were segmented based on the standardized images, and the visualization and manual correction were performed; the shape features, first-order statistical features, texture features and higher-order features were extracted using radiomics software packages, and the features were screened and dimensionality reduced; the Radscore model was constructed based on the screened radiomics features, and logistic regression was used for training and cross-validation. (4) Construction of a joint clinical feature and radiomics model: The output of the Radscore model is combined with the selected clinical features; a joint prediction model is constructed and trained and validated by logistic regression to achieve prediction of different types of temporal lobe epilepsy; (5) Model performance comparison and evaluation: The performance of clinical models, radiomics models and combined models is evaluated; statistical comparison and validation of models are performed to quantify the performance differences and potential clinical application value between different models; (6) Visualization of prediction results and individualized prediction: Based on the joint model, a nomogram for the diagnosis and prediction of different types of temporal lobe epilepsy is constructed; the prediction results are visualized and individualized risk prediction is provided for each patient to assist clinicians in making early subtype diagnosis and treatment decisions.

2. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, The clinical data mentioned in step (1) includes demographic information, disease course, epileptic seizure type and frequency, comorbid history, and cognitive-emotional scale scores.

3. The method for predicting different types of temporal lobe epilepsy based on MRI radiomics and clinical characteristics according to claim 1, characterized in that, The MRI images mentioned in step (1) are the brain MRI images of the patient after admission, including axial 3D-T1 weighted images, T2 weighted images, and T2 fluid attenuation inversion recovery sequences, which are exported in DICOM format for subsequent image processing.

4. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, Step (2) describes the process of predicting potential clinical factors for different types of temporal lobe epilepsy as follows: For each clinical feature X i A univariate logistic regression model was constructed to predict the type Y of temporal lobe epilepsy. For the screened potential clinical factors X i1 ,X i2 ,…,X ik Perform multivariate logistic regression analysis to establish a multivariate clinical characteristic prediction model; Where P(Y=1) represents the probability of developing a certain type of temporal lobe epilepsy; β0 is the intercept term; β1,...,β k represents the regression coefficients for each clinical feature.

5. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, The preprocessing described in step (3) includes motion correction, non-brain tissue stripping, intensity normalization, and spatial registration.

6. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, The key temporal lobe structure mentioned in step (3) is the bilateral hippocampus; the radiomics software package is pyRadiomics; the feature selection and dimensionality reduction include K-optimal, recursive feature elimination, and the LASSO method. For the K-optimal method, the k features with the highest correlation to category Y (epilepsy type) are selected: For LASSO dimensionality reduction, key features are selected using L1 regularization: Where θ(·) is the loss function and λ is the regularization coefficient.

7. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, The performance evaluation in step (4) includes calculating the subject operating characteristic curve and area under the curve, sensitivity, specificity, accuracy, and F1 score.

8. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, In the column chart described in step (6), the first row is the score scale, with a score range of 0 to 100; the second row is the age, with a score range of 15 to 85; the third row is the Radscore value, with a score range of -1.6 to 1.4; and the fourth row is the total score, with a score range of 0 to 120. The fifth row contains linear prediction values, ranging from -9 to 5; the sixth row contains prediction probabilities, ranging from 0.1 to 0.

9.

9. The method for establishing a nomogram model for the diagnosis and prediction of different types of temporal lobe epilepsy according to claim 1, characterized in that, In the nomogram described in step (6), the first row is the score scale, with a score range of 0 to 100; the second row is the epilepsy type, with a score range of 0 to 1; the third row is the Radscore value, with a score range of -1.4 to 1 point; and the fourth row is the total score, with a score range of 0 to 120 points. The fifth row contains linear prediction values, ranging from -7 to 7; the sixth row contains prediction probabilities, ranging from 0.1 to 0.9.

Images