Parkinson's disease (PD) is the second most common neurodegenerative disorder. PD usually arises with neurons' death, thus leading to the decrease of the amount of dopamine they produce in the brain. Hence, as the quantity of dopamine reduces, PD starts to take place very gradually. Its primary symptoms are tremor, slow movements, muscle stiffness, unstable posture, and loss of balance. For some patients, a lot of symptoms can be refractory to almost every medication and this is when they undergo deep brain stimulation (DBS). DBS consists of delivering continuous electrical stimulation through chronically implanted electrodes to a neural brain target structure. The subthalamic nucleus (STN) is a very popular target region. Since it is very small, it is poorly visualized on conventional imaging and the microelectrode recording (MER) is commonly used to refine STN targeting. In this work, we have investigated MER-based features to characterize the functional activity recorded during electrophysiological exploration and to identify the STN based on classification models. MERs from thirteen PD patients were obtained at different depths. For each patient three parallel MER tracks (anterior, central and posterior) were performed in both the hemispheres with 1 or 0.5 mm steps, from 10mm above to 5mm beyond the estimated target along the electrode trajectory. Recordings lasted 10 s for each depth. After removing visually identified large artifacts, spike detection was performend. Then, we implemented a semiautomatic spike sorting algorithm. Of note, not every cluster created will represent a single unit. We distinguished well defined clusters that might be single units from ones that are noise. We then extracted a total of 23 features from different domains: time, frequency, action potentials and single units. We evaluated features usefulness to distinguish the STN from other structures crossed along the electrode trajectory, i.e. thalamus (TH), zona incerta (ZI), and substantia nigra (SN). Multiple classification models were feeded with both the original features set and the different created features subsets allowing to recognize the STN location depth along the explored electrode trajectory. A classification was obtained by two independent experts so as to make the clinical classification as robust and reliable as possible and to obtain an adequate ground truth. The presented features subsets allowed to reach mean validation test accuracies in the range [82.78% 93.41%] through the various trained classifiers. Frequency domain features subset showed one of the highest accuracy value (92.25%), suggesting how the signal total power could be a good indicator for the STN depth identification. This statement is confirmed by observing how its trend along the explored depths strongly increases when the STN area is crossed. In conclusion, in this work we explored different MER-based features sets for the identification of an optimal combination able to discriminate STN and NON-STN recordings in a machine-learning framework. Our finding suggested that good results can be obtained on our dataset, even with a single feature based on power spectral density analysis.
Il morbo Parkinson è una malattia neuro-degenerativa derivante dalla progressiva deprivazione della dopamina, responsabile del controllo del movimento. Instabilità posturale, bradicinesia, tremori a riposo e rigidità muscolare sono i sintomi principali del PD, alleviati attualmente con diverse tipologie di trattamento, farmacologiche e non. Quando i farmaci perdono l’efficacia vengono adottate alternative chirurgiche. Tra queste la stimolazione cerebrale profonda (DBS), che consiste nell'invio di impulsi elettrici ad alta frequenza in aree cerebrali specifiche. Il bersaglio più frequentemente scelto per il trattamento è il nucleo subtalamico (STN), la cui iperattività sostiene i sintomi disabilitanti caratteristici della malattia. Poiché l'STN è una struttura molto piccola, le tecniche di imaging convenzionali da sole non sono sufficienti ad individuarlo e sono di norma integrate con la registrazione con microelettrodi (MER). In questo lavoro, abbiamo analizzato un gruppo di caratteristiche estratte da segnali MER per differenziare l'attività funzionale registrata durante l'esplorazione elettrofisiologica ed identificare l’STN tramite modelli di classificazione. Sono stati acquisiti segnali MER da tredici pazienti con Parkinson. Per ciascun paziente sono state eseguite tre tracce MER parallele (anteriore, centrale e posteriore) in entrambi gli emisferi con incrementi di 1 o 0,5 mm, da 10 mm sopra a 5 mm oltre il target chirurgico stimato lungo la traiettoria dell'elettrodo ed effettuando una registrazione di 10 s per ogni profondità. Dopo aver rimosso evidenti artefatti, abbiamo rilevato gli spike presenti. Quindi, abbiamo implementato un algoritmo di spike sorting semi-automatico così da assegnare gli spike trovati ai diversi neuroni presenti nell'area esplorata. In particolare, sono stati distinti i cluster rappresentanti le singole unità da quelli rumorosi. Sono state estratte 23 caratteristiche appartenenti a diversi domini: tempo, frequenza, potenziali d'azione e singole unità. Diversi modelli di classificazione sono stati alimentati sia con il dataset di caratteristiche originale sia con singoli sottogruppi di caratteristiche selezionati permettendo di riconoscere la profondità corrispondente alla struttura target lungo la traiettoria dell'elettrodo. I risultati ottenuti sono stati confrontati con la classificazione eseguita da due esperti indipendenti in modo da rendere quanto più robusta e affidabile possibile la classificazione clinica. Le caratteristiche selezionate hanno permesso di ottenere dei valori di accuratezza di validazione media tra [82.78% 93.41%] durante la fase di test dei classificatori allenati. In particolare, la potenza totale del segnale, unica caratteristica del dominio della frequenza, ha permesso di ottenere uno dei più alti valori di accuratezza media (92.25%), suggerendo come la stessa possa essere un buon indicatore, di supporto al clinico, nella localizzazione del target neurale. Questo è stato confermato osservando l'andamento della stessa in funzione della distanza dal target: la potenza totale, infatti, aumenta notevolmente quando viene attraversato l'STN. In conclusione, in questo lavoro abbiamo esplorato diversi set di caratteristiche basate su MER per identificare una combinazione ottimale in grado di discriminare le registrazioni STN e NON-STN in un framework di apprendimento automatico. I risultati ottenuti suggeriscono che è possibile ottenere buoni risultati nel nostro set di dati, anche con una singola caratteristica basata sull'analisi della densità spettrale di potenza.
Evaluation of microelectrode recording-based features for subthalamic nucleus identification in deep brain stimulation surgery
DEL VECCHIO DEL VECCHIO, JASMIN
2019/2020
Abstract
Parkinson's disease (PD) is the second most common neurodegenerative disorder. PD usually arises with neurons' death, thus leading to the decrease of the amount of dopamine they produce in the brain. Hence, as the quantity of dopamine reduces, PD starts to take place very gradually. Its primary symptoms are tremor, slow movements, muscle stiffness, unstable posture, and loss of balance. For some patients, a lot of symptoms can be refractory to almost every medication and this is when they undergo deep brain stimulation (DBS). DBS consists of delivering continuous electrical stimulation through chronically implanted electrodes to a neural brain target structure. The subthalamic nucleus (STN) is a very popular target region. Since it is very small, it is poorly visualized on conventional imaging and the microelectrode recording (MER) is commonly used to refine STN targeting. In this work, we have investigated MER-based features to characterize the functional activity recorded during electrophysiological exploration and to identify the STN based on classification models. MERs from thirteen PD patients were obtained at different depths. For each patient three parallel MER tracks (anterior, central and posterior) were performed in both the hemispheres with 1 or 0.5 mm steps, from 10mm above to 5mm beyond the estimated target along the electrode trajectory. Recordings lasted 10 s for each depth. After removing visually identified large artifacts, spike detection was performend. Then, we implemented a semiautomatic spike sorting algorithm. Of note, not every cluster created will represent a single unit. We distinguished well defined clusters that might be single units from ones that are noise. We then extracted a total of 23 features from different domains: time, frequency, action potentials and single units. We evaluated features usefulness to distinguish the STN from other structures crossed along the electrode trajectory, i.e. thalamus (TH), zona incerta (ZI), and substantia nigra (SN). Multiple classification models were feeded with both the original features set and the different created features subsets allowing to recognize the STN location depth along the explored electrode trajectory. A classification was obtained by two independent experts so as to make the clinical classification as robust and reliable as possible and to obtain an adequate ground truth. The presented features subsets allowed to reach mean validation test accuracies in the range [82.78% 93.41%] through the various trained classifiers. Frequency domain features subset showed one of the highest accuracy value (92.25%), suggesting how the signal total power could be a good indicator for the STN depth identification. This statement is confirmed by observing how its trend along the explored depths strongly increases when the STN area is crossed. In conclusion, in this work we explored different MER-based features sets for the identification of an optimal combination able to discriminate STN and NON-STN recordings in a machine-learning framework. Our finding suggested that good results can be obtained on our dataset, even with a single feature based on power spectral density analysis.File | Dimensione | Formato | |
---|---|---|---|
2020_4_DelVecchioDelVecchio.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
5.41 MB
Formato
Adobe PDF
|
5.41 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/10589/154562