The study of sleep-related disorders has become increasingly important due to their negative impact on overall health. This thesis was conducted in collaboration with Biocubica S.r.l., which has developed a wearable device called Soundi. The Soundi is capable of acquiring various surrogate signals in a multimodal mode and can replace traditional polysomnography. The device was used in a study conducted at Auxologico 'San Luca' Hospital in Milan on 50 patients. The aim was to compare the signals measured by the Soundi with those obtained by PSG. The goals of this thesis are twofold: to create a cloud structure capable of collecting and processing the examinations of the study patients and to develop an AI network for the automatic recognition of sleep apnea. The cloud operations involve decompressing the DAT file in the Soundi, resampling it at 400 Hz, and processing the exam in EDF format for physician labeling. An application was used to enable communication with the cloud, allowing the physician to schedule the exam date and time, trigger the Soundi, upload the recorded file to the cloud, and download it in processed EDF format. For the AI part, a dataset was created with the data collected from the patients, in which each second was labeled with a specific target, including Central Apnea, Mixed Apnea, Obstructive Apnea, Artifact, Movement, Hypoapnea, and Normal. Time series of 10-second duration were subsequently created. To address the AI component, a dataset was generated using patient data. To prevent dataset imbalance, an under-sampling method was applied to a portion of the data. For analysis, we selected a CNN network and divided the dataset into training (70%), validation (15%), and test (15%) sets. The results obtained showed a good level of network learning, also demonstrated by the gradCAM graphical display method, which shows worse results for central and mixed apneas, since they contain less data than the others.
L’analisi dei disturbi legati al sonno ha acquisito rilevanza dato che questi influenzano negativamente la salute generale dell’individuo. La tesi è stata realizzata in collaborazione con Biocubica S.r.l. che ha sviluppato un dispositivo indossabile, il Soundi. Questo device è in grado di acquisire diversi segnali surrogati in modalità multimodale e può essere utilizzato per sostituire la tradizionale polisonnografia. Il dispositivo è stato utilizzato all’interno del mio studio con l’Ospedale Auxologico ‘San Luca’ di Milano su 50 pazienti, con lo scopo di confrontare i segnali misurati dal Soundi con quelli ottenuti dalla PSG. Gli obiettivi di questa tesi sono due: creare una struttura cloud in grado di raccogliere ed elaborare gli esami dei pazienti dello studio e sviluppare una rete di AI per il riconoscimento automatico delle apnee notturne. Le operazioni eseguite all’interno del cloud consistono nel decomprimere il file DAT presente nel Soundi, ricampionare il file a 400 Hz ed elaborare l’esame in formato EDF in modo che possa essere etichettato dal medico. Per permettere la comunicazione con il cloud, è stata utilizzata un’applicazione che consentisse al medico sia di programmare la data e l’orario dell’esame per l’attivazione del Soundi sia di caricare il file registrato dal Soundi sul cloud per poi scaricarlo elaborato in formato EDF. Per la parte di AI, è stato creato un dataset con i dati raccolti dai pazienti, in cui è stato etichettato ogni secondo con un target specifico (Apnea Centrale, Apnea Mista, Apnea Ostruttiva, Artefatto, Movimento, Ipoapnea, Normale), poi sono state create delle time series di 10 secondi di durata. Per evitare lo sbilanciamento del dataset, è stato applicato un metodo di under-sampling su una parte dei dati. Una rete CNN è stata scelta per l’analisi e il dataset è stato diviso in training (70%), validation (15%) e test (15%) set. I risultati ottenuti hanno evidenziato un buon livello di apprendimento della rete, dimostrato anche dal metodo di visualizzazione grafica della gradCAM, che mostra risultati peggiori per le apnee centrali e miste, dato che contengono meno dati rispetto alle altre.
Cloud management infrastructure to collect polysomnographic data with wearable device in order to detect sleep apnea events using AI
CIUFFREDA, MATTEO
2023/2024
Abstract
The study of sleep-related disorders has become increasingly important due to their negative impact on overall health. This thesis was conducted in collaboration with Biocubica S.r.l., which has developed a wearable device called Soundi. The Soundi is capable of acquiring various surrogate signals in a multimodal mode and can replace traditional polysomnography. The device was used in a study conducted at Auxologico 'San Luca' Hospital in Milan on 50 patients. The aim was to compare the signals measured by the Soundi with those obtained by PSG. The goals of this thesis are twofold: to create a cloud structure capable of collecting and processing the examinations of the study patients and to develop an AI network for the automatic recognition of sleep apnea. The cloud operations involve decompressing the DAT file in the Soundi, resampling it at 400 Hz, and processing the exam in EDF format for physician labeling. An application was used to enable communication with the cloud, allowing the physician to schedule the exam date and time, trigger the Soundi, upload the recorded file to the cloud, and download it in processed EDF format. For the AI part, a dataset was created with the data collected from the patients, in which each second was labeled with a specific target, including Central Apnea, Mixed Apnea, Obstructive Apnea, Artifact, Movement, Hypoapnea, and Normal. Time series of 10-second duration were subsequently created. To address the AI component, a dataset was generated using patient data. To prevent dataset imbalance, an under-sampling method was applied to a portion of the data. For analysis, we selected a CNN network and divided the dataset into training (70%), validation (15%), and test (15%) sets. The results obtained showed a good level of network learning, also demonstrated by the gradCAM graphical display method, which shows worse results for central and mixed apneas, since they contain less data than the others.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Ciuffreda_Tesi_01.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
6.73 MB
Formato
Adobe PDF
|
6.73 MB | Adobe PDF | Visualizza/Apri |
2024_04_Ciuffreda_Executive Summary_02.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
504.91 kB
Formato
Adobe PDF
|
504.91 kB | 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/10589/218908