Photoplethysmography (PPG) is an effective technique to measure blood oxygen levels (SpO2) and blood perfusion in peripheral circulation, used estimate pulse rate (PR). Despite limited high-quality validation studies, PPG remains promising, especially for wearable devices. PPG signals are typically recorded at the finger for higher quality compared to signals from other body locations, such as the wrist or chest. A generative model is proposed to transform noisy chest PPG (cPPG) signals into high-quality finger PPG (fPPG) signals using a style transfer-assisted cycle coherent generative adversarial network (stccGAN). PPG data was collected from 30 subjects using two identical devices capable of recording red, infrared, and green wavelengths. Before feeding them to the model, both acquired cPPG and acquired fPPG signals were filtered using a bidirectional 4th order band pass Bessel filter (0.5-15 Hz), segmented into overlapping 5 second windows, standardized with the Z-score, and then split into train, validation, and test sets, in a subject-wise manner. Eleven cycle GANs configurations were developed with different complexity levels. Performance was evaluated using RRMSE in both frequency and time domains; and MAE, R, R2, and clinical indices such as peak to peak (PP) intervals and PR in the time domain only. The proposed model achieved significant improvements from the same metrics initially computed between the acquired cPPG and the reference fPPG. Respectively, the metrics improved of: 58.32% for the frequency RRMSE, 42.65% for the time RRMSE, 47.78% for the MAE, and 141.34% for R. R2 already indicates a percentage which represents the amount of variance explained by the prediction of the model, which achieved an average absolute value of 59.00%, from an initial amount close to 0%. For the additional evaluation indices, the Pearson correlation coefficient (PCC) between predicted and actual PP intervals and PR values averaged, respectively, 0.55 and 0.63, demonstrating overall good alignment with fPPG derived PP interval and PR values.
La fotopletismografia (PPG) è una tecnica efficace per misurare i livelli di ossigeno nel sangue e la perfusione sanguigna nella circolazione periferica, usata per la stima della frequenza cardiaca (PR). Nonostante accurati studi di validazione siano ancora limitati, il segnale PPG si conferma una tecnologia promettente, soprattutto nel campo dei dispositivi indossabili. I segnali PPG vengono generalmente acquisiti dal dito (fPPG) per ottenere dati di qualità superiore rispetto a quelli raccolti da altre aree del corpo, come ad esempio il polso o il torace (cPPG). In questo lavoro è stato proposto un modello generativo per trasformare i segnali rumorosi di cPPG in segnali di alta qualità di fPPG, utilizzando una rete avversaria generativa coerente ciclica assistita da trasferimento di stile (stccGAN). I dati PPG sono stati raccolti da 30 soggetti differenti utilizzando due dispositivi identici capaci di registrare il segnale PPG attraverso 3 lunghezze d’onda: rosse, infrarosse e verdi. I segnali cPPG e fPPG acquisiti sono stati filtrati utilizzando un filtro passa-banda bidirezionale di Bessel del 4° ordine (0.5-15 Hz), segmentati in finestre di 5 secondi parzialmente sovrapposte, standardizzati tramite Z-score e suddivisi in training set, validation set e test set, in modo soggetto-specifico. Sono state sviluppate undici configurazioni di cycle GAN con livelli di complessità differenti. Le prestazioni sono state valutate utilizzando metriche come RRMSE nel dominio della frequenza e nel dominio del tempo; MAE, R, R2 nel dominio del tempo. In aggiunta, sono stati calcolati anche degli indici, quali l’intervall picco-picco (PP) e la PR. Il modello proposto ha portato a miglioramenti significativi nei seguenti parametri rispetto ai valori ottenuti sulle stesse metriche confrontando il cPPG acquisito e l’fPPG di riferimento. Rispettivamente, le metriche sono migliorate del: 58,32% per l’RRMSE nel dominio delle frequenze, 42,65% per l’RRMSE nel dominio del tempo, 47,78% per la MAE, e del 141,34% per R. R2 indica già un valore percentuale che rappresenta la varianza spiegata dal modello che ha raggiunto un valore assoluto medio tra i tre canali del 59.00%, partendo da un valore iniziale vicino allo 0%. Per quanto riguarda gli indici, la correlazione di Pearson tra i valori PP e PR predetti dal modello proposto e i valori PP e PR effettivi, misurati sull’fPPG di riferimento; ha ottenuto un valore medio rispettivamente di 0.55 per gli intervalli PP e di 0,63 per PR, dimostrando un buon allineamento con i rispettivi valori derivati dal fPPG.
Style transfer-assisted deep learning method for photplethysmogram denoising
TOGNONI, FEDERICO
2023/2024
Abstract
Photoplethysmography (PPG) is an effective technique to measure blood oxygen levels (SpO2) and blood perfusion in peripheral circulation, used estimate pulse rate (PR). Despite limited high-quality validation studies, PPG remains promising, especially for wearable devices. PPG signals are typically recorded at the finger for higher quality compared to signals from other body locations, such as the wrist or chest. A generative model is proposed to transform noisy chest PPG (cPPG) signals into high-quality finger PPG (fPPG) signals using a style transfer-assisted cycle coherent generative adversarial network (stccGAN). PPG data was collected from 30 subjects using two identical devices capable of recording red, infrared, and green wavelengths. Before feeding them to the model, both acquired cPPG and acquired fPPG signals were filtered using a bidirectional 4th order band pass Bessel filter (0.5-15 Hz), segmented into overlapping 5 second windows, standardized with the Z-score, and then split into train, validation, and test sets, in a subject-wise manner. Eleven cycle GANs configurations were developed with different complexity levels. Performance was evaluated using RRMSE in both frequency and time domains; and MAE, R, R2, and clinical indices such as peak to peak (PP) intervals and PR in the time domain only. The proposed model achieved significant improvements from the same metrics initially computed between the acquired cPPG and the reference fPPG. Respectively, the metrics improved of: 58.32% for the frequency RRMSE, 42.65% for the time RRMSE, 47.78% for the MAE, and 141.34% for R. R2 already indicates a percentage which represents the amount of variance explained by the prediction of the model, which achieved an average absolute value of 59.00%, from an initial amount close to 0%. For the additional evaluation indices, the Pearson correlation coefficient (PCC) between predicted and actual PP intervals and PR values averaged, respectively, 0.55 and 0.63, demonstrating overall good alignment with fPPG derived PP interval and PR values.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Tognoni_Tesi.pdf
non accessibile
Descrizione: Testo della Tesi
Dimensione
16.03 MB
Formato
Adobe PDF
|
16.03 MB | Adobe PDF | Visualizza/Apri |
2024_12_Tognoni_Executive_Summary.pdf
solo utenti autorizzati a partire dal 19/11/2025
Descrizione: Testo Executive Summary
Dimensione
1.84 MB
Formato
Adobe PDF
|
1.84 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/10589/230888