Human stereo-encephalography (SEEG) acquisitions are exponentially growing, and their interest in the field of human neuroscience is steadily increasing through the years. Being a method that was born in clinical scenarios for the treatment of drug-resistant epileptic patients,SEEG allows to pursue study in the field of interactions within and across brain networks for functional mapping of the brain with high spatial and temporal resolution, and high signal to noise ratio (SNR). However, such data are characterized by a plethora of artifacts spacing from electrical artifacts to epileptic waveforms due to the pathological condition, making manual cleaning a mandatory step in every SEEG study. Considering these premises, an automatic method to handle these huge amounts of data is highly desired. Thus, in the present thesis a possible approach to deal with this specific challenge is presented. With the aim of providing automaticity in the process, a gradient boosting machine learning model was developed. The model was trained by using the SEEG channels from nine drug-resistant epileptic patients; from every channel simple features were extracted, spanning across the time and frequency domain. The model achieved an overall precision of 80%, recall 55%, and accuracy 91%. Moreover, to provide a semi-automatic tool for the cleaning of the SEEG signal, a graphical user interface (GUI) was developed. The GUI exploits the developed model in order to reject (or accept) channels, while the epochs are rejected with custom functionalities. Moreover, there is the possibility to implement a custom-made rejection algorithm allowing the tool to be highly customizable. Overall, with the provided model (gradient boosting model) and GUI, it was possible to provide high precision and faster cleaning time, in respect to the usual manual cleaning.
Nel corso degli ultimi anni, le acquisizioni di stereo-encefalografia umana (SEEG) stanno crescendo in modo esponenziale e il loro interesse nel campo delle neuroscienze umane è in costante aumento. Essendo un metodo nato in ambiti prettamente clinici per il trattamento di pazienti che presentano un'epilessia farmacoresistente, la SEEG permette di proseguire studi nel campo delle interazioni tra le reti cerebrali per la mappatura funzionale del cervello, garantendo alta risoluzione sia spaziale che temporale, con un rapporto segnale-rumore favorevole (SNR). Tuttavia, i dati ricavata da questa modalità sono caratterizzati da una numerose tipologie di artefatti spaziando da artefatti elettrici arrivando alla presenza di forme d'onda epilettiche dovute alle condizioni patologiche dei pazienti da cui questi dati vengono misurati, rendendo la pulizia manuale un passaggio obbligatorio in ogni studio SEEG. Considerando queste premesse, è fortemente auspicabile un metodo automatico per gestire queste enormi quantità di dati. Pertanto, nella presente tesi viene presentato un possibile approccio per affrontare questo problema. È stato sviluppato un modello di apprendimento automatico basato su un modello “gradient boosting” con apprendimento supervisionato al fine di fornire dell’automatismo nella pulizia dei dati. Il modello è stato allenato utilizzando i canali SEEG proveniente da nove pazienti con epilessia farmacoresistente; da ogni canale sono state estratte delle semplici features, nel dominio del tempo e della frequenza. Il modello ha raggiunto una precisione complessiva dell'80%, un recall del 55% e un'accuratezza del 91%. Inoltre, è stata sviluppata una interfaccia grafica per l’utente (GUI) per fornire uno strumento che possa portare ad una pulizia semiautomatica del segnale SEEG. La GUI sfrutta il modello sviluppato, o eventualmente altri algoritmi, per rifiutare (o accettare) i canali, mentre le epoche vengono rifiutate con ulteriori funzionalità. Inoltre, lo strumento permette l’implementazione di algoritmi custom-made di rimozione di canali, garantendogli di essere altamente personalizzabile. Nel complesso, con il modello fornito (modello gradient boosting) e la GUI, è stato possibile fornire un'elevata precisione e tempi di pulizia più rapidi, rispetto alla consueta pulizia manuale.
A novel tool for semi-automatic SEEG bad channels and epochs rejection to improve the cleaning procedure of SPES protocol analysis
Murari, Alessandro
2020/2021
Abstract
Human stereo-encephalography (SEEG) acquisitions are exponentially growing, and their interest in the field of human neuroscience is steadily increasing through the years. Being a method that was born in clinical scenarios for the treatment of drug-resistant epileptic patients,SEEG allows to pursue study in the field of interactions within and across brain networks for functional mapping of the brain with high spatial and temporal resolution, and high signal to noise ratio (SNR). However, such data are characterized by a plethora of artifacts spacing from electrical artifacts to epileptic waveforms due to the pathological condition, making manual cleaning a mandatory step in every SEEG study. Considering these premises, an automatic method to handle these huge amounts of data is highly desired. Thus, in the present thesis a possible approach to deal with this specific challenge is presented. With the aim of providing automaticity in the process, a gradient boosting machine learning model was developed. The model was trained by using the SEEG channels from nine drug-resistant epileptic patients; from every channel simple features were extracted, spanning across the time and frequency domain. The model achieved an overall precision of 80%, recall 55%, and accuracy 91%. Moreover, to provide a semi-automatic tool for the cleaning of the SEEG signal, a graphical user interface (GUI) was developed. The GUI exploits the developed model in order to reject (or accept) channels, while the epochs are rejected with custom functionalities. Moreover, there is the possibility to implement a custom-made rejection algorithm allowing the tool to be highly customizable. Overall, with the provided model (gradient boosting model) and GUI, it was possible to provide high precision and faster cleaning time, in respect to the usual manual cleaning.File | Dimensione | Formato | |
---|---|---|---|
Thesis Alessandro.pdf
non accessibile
Dimensione
21.66 MB
Formato
Adobe PDF
|
21.66 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/10589/182694