Temporal lobe epilepsy (TLE) is the most common drug-resistant form of epilepsy. TLE is increasingly understood not merely as a focal disorder, but as a large-scale network disease characterized by whole-brain alterations in connectivity. Using the Epilepsy Connectome Project (ECP) dataset (105 TLE, 76 HC), this thesis aims to optimally design the feature space for TLE classification by leveraging the significant differences in structural connectivity between TLE patients and healthy controls. Our approach integrated conventional microstructural di!usion metrics with advanced network-based metrics derived from graph theory to quantify brain reorganization both locally and globally. The novelty of this work lies in the introduction of a multi-scale and multi-domain framework for feature extraction, which accounts for both anatomical and functional brain networks, such as the Limbic System and the Default Mode Network (DMN), notably the two most affected by TLE. By doing so, this work bridges the gap between advanced structural connectivity analysis and machine learning-based clinical classification. Specifically, we developed a Machine Learning classification pipeline integrating the non- parametric Mann-Whitney U test and Spearman correlation for informed feature selection. Our final Support Vector Classifier (SVC) model, which combines local diffusion properties and global network topology derived from the hierarchical framework, reached mean cross-validation values of 84.4% for ROC-AUC, 75.3% for Balanced Accuracy and 78.6% for Recall, maintaining consistent performance on the independent test set and thus demonstrating the model’s generalizability. These results confirm the efficacy of combining diffusion metrics and connectomics and demonstrate that feature design based on a multi-level hierarchical framework can effectively capture the complex pathological signatures of TLE.
L’epilessia del lobo temporale (TLE) è la forma più comune di epilessia farmaco-resistente. La TLE è oggi considerata non solo un disturbo focale, ma una patologia di rete, caratterizzata da alterazioni della connettività dell’intero encefalo. Utilizzando il dataset dell’Epilepsy Connectome Project (ECP) (105 TLE, 76 HC), questa tesi si propone di analizzare le differenze significative riguardanti la connettività strutturale tra pazienti con TLE e controlli sani, al fine di progettare in modo ottimale lo spazio delle variabili per la classificazione della patologia. Il nostro approccio integra metriche di diffusione microstrutturale convenzionali con metriche di rete derivate dalla teoria dei grafi, per quantificare la riorganizzazione cerebrale sia a livello locale che globale. La novità di questo lavoro risiede nell’introduzione di un framework multiscala e multidominio per l’estrazione dei predittori, in grado di modellare sia le dinamiche delle reti cerebrali anatomiche che di quelle funzionali. Tra esse, il Sistema Limbico e la Default Mode Network (DMN) sono considerati i due sistemi più colpiti dalla TLE. Dunque, questo lavoro colma il divario tra l’analisi avanzata della connettività strutturale nella TLE e la classificazione clinica basata sul machine learning. Nello specifico, abbiamo sviluppato una pipeline di classificazione basata sul Machine Learning che integra il test non-parametrico di Mann-Whitney e la correlazione di Spearman per la selezione delle feature. Il nostro modello finale, un Support Vector Classifier (SVC) che combina proprietà di di!usione locale e topologia di rete globale derivate dal framework gerarchico, ha raggiunto valori medi in cross-validation dell’84,4% per la ROC-AUC, del 75,3% per l’Accuratezza Bilanciata e del 78,6% per la Sensitività e ha mantenuto prestazioni coerenti sul test set indipendente, dimostrando così la generalizzabilità del modello. Infine, questi risultati supportano l’efficacia della combinazione di metriche di diffusione e connettomica e dimostrano che l’utilizzo dell’approccio a più livelli gerarchici anatomici e funzionali per la costruzione dello spazio delle variabili può catturare efficacemente le complesse dinamiche della TLE.
A New Feature Extraction Approach for Temporal Lobe Epilepsy Classification: Integrating Conventional Diffusion and Structural Network Metrics in a Multi-Scale Framework
BROSERA', ELISA
2025/2026
Abstract
Temporal lobe epilepsy (TLE) is the most common drug-resistant form of epilepsy. TLE is increasingly understood not merely as a focal disorder, but as a large-scale network disease characterized by whole-brain alterations in connectivity. Using the Epilepsy Connectome Project (ECP) dataset (105 TLE, 76 HC), this thesis aims to optimally design the feature space for TLE classification by leveraging the significant differences in structural connectivity between TLE patients and healthy controls. Our approach integrated conventional microstructural di!usion metrics with advanced network-based metrics derived from graph theory to quantify brain reorganization both locally and globally. The novelty of this work lies in the introduction of a multi-scale and multi-domain framework for feature extraction, which accounts for both anatomical and functional brain networks, such as the Limbic System and the Default Mode Network (DMN), notably the two most affected by TLE. By doing so, this work bridges the gap between advanced structural connectivity analysis and machine learning-based clinical classification. Specifically, we developed a Machine Learning classification pipeline integrating the non- parametric Mann-Whitney U test and Spearman correlation for informed feature selection. Our final Support Vector Classifier (SVC) model, which combines local diffusion properties and global network topology derived from the hierarchical framework, reached mean cross-validation values of 84.4% for ROC-AUC, 75.3% for Balanced Accuracy and 78.6% for Recall, maintaining consistent performance on the independent test set and thus demonstrating the model’s generalizability. These results confirm the efficacy of combining diffusion metrics and connectomics and demonstrate that feature design based on a multi-level hierarchical framework can effectively capture the complex pathological signatures of TLE.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_3_Brosera_ExecutiveSummary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary Elisa Broserà
Dimensione
8.21 MB
Formato
Adobe PDF
|
8.21 MB | Adobe PDF | Visualizza/Apri |
|
2026_3_Brosera_Tesi.pdf
accessibile in internet per tutti
Descrizione: Tesi Elisa Broserà
Dimensione
24.01 MB
Formato
Adobe PDF
|
24.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/252540