Pregnancy necessitates a large number of physiological changes in the mother. These changes are mediated by the autonomic nervous system (ANS) function. In fact, ANS imbalance is related to pregnancy complications. Given the percentage of pregnant women who develop complications, it is important to find a screening tool that can identify women at risk and monitor their health status during pregnancy. Photoplethysmography (PPG) could be a solution to this need, as it allows the analysis of heart rate variability (HRV), which is closely related to ANS function. In addition, PPG makes it possible to determine features that describe vascular function, which is of interest in the context of the study of the physiology of pregnancy. To understand the potential of PPG as a screening tool in obstetrics and to monitor at-risk pregnancies, it is important to understand the differences that occur in healthy pregnancies. This study aims to identify whether differences in HRV and PPG morphology between pregnant and non-pregnant women are significant and large enough to discriminate these populations using a machine learning approach. High performance is desired, but the core purpose of the model is to provide insights into features useful for this discrimination. To fulfill the objectives, several data transformation steps were considered to improve the performance of the classifiers. In addition, multiple classifiers were implemented. A pipeline structure was used to identify the best combination of transformations and classifiers. The pipelines input were the features computed from HRV and PPG morphology. A pipeline using logistic regression as classifier was selected as the best model for this application because it represented a good compromise between performance and interpretability. The selected model performed well using ten parameters, ordered by the weights assigned by the logistic regression function. The model provided insights into which features are relevant to this classification and to what extent. These results emphasized the importance of using HRV features that are not commonly included in studies and the added value of combining the study of HRV with PPG morphology features.
La gravidanza richiede un gran numero di cambiamenti fisiologici nella madre. Questi cambiamenti sono mediati dalla funzione del sistema nervoso autonomo (ANS). Infatti, lo squilibrio dell'ANS è legato alle complicazioni della gravidanza. Date queste complicazioni, è importante trovare uno strumento di screening per identificare le donne a rischio e monitorarne lo stato di salute. La fotopletismografia (PPG) potrebbe rappresentare una soluzione a questa esigenza, grazie all'analisi della variabilità della frequenza cardiaca (HRV), che è strettamente correlata alla funzione dell’ANS. Inoltre, la PPG permette di determinare le caratteristiche che descrivono la funzione vascolare. Per comprendere il potenziale della PPG come strumento di screening, è importante capire le differenze che si presentano in caso di gravidanza sana. Questo studio si propone di identificare se le differenze nella HRV e nella morfologia del PPG tra donne incinte e non incinte sono sufficientemente rilevanti per la discriminazione di queste popolazioni con un approccio di machine learning. È desiderato ottenere un rendimento elevato, ma lo scopo principale del modello è quello di fornire indicazioni sulle caratteristiche utili per la discriminazione. Per raggiungere gli obiettivi, sono state prese in considerazione diverse fasi di trasformazione dei dati per potenziare la performance dei classificatori. Inoltre, sono stati implementati diversi classificatori. Per questo è stata utilizzata una struttura a pipeline. L’input della pipeline era costituito dalle caratteristiche calcolate dalla morfologia di HRV e PPG. Come modello migliore per questa applicazione è stata scelta una pipeline che utilizza la regressione logistica come classificatore, in quanto rappresenta un buon trade-off tra performance e interpretabilità. Il modello selezionato ha ottenuto un buon livello di performance utilizzando 10 caratteristiche. Il modello ha fornito indicazioni sulle caratteristiche che sono rilevanti per questa classificazione e in quale misura. Ciò, ha sottolineato l’importanza di utilizzare feature dell'HRV che non sono comunemente incluse negli studi e il valore aggiunto di combinare lo studio dell’HRV con le caratteristiche della morfologia del PPG.
Classification of pregnant and non-pregnant women based on heart rate variability and PPG pulse wave morphology features - A machine learning approach.
ALMARIO ESCORCIA, MARÍA JOSÉ
2021/2022
Abstract
Pregnancy necessitates a large number of physiological changes in the mother. These changes are mediated by the autonomic nervous system (ANS) function. In fact, ANS imbalance is related to pregnancy complications. Given the percentage of pregnant women who develop complications, it is important to find a screening tool that can identify women at risk and monitor their health status during pregnancy. Photoplethysmography (PPG) could be a solution to this need, as it allows the analysis of heart rate variability (HRV), which is closely related to ANS function. In addition, PPG makes it possible to determine features that describe vascular function, which is of interest in the context of the study of the physiology of pregnancy. To understand the potential of PPG as a screening tool in obstetrics and to monitor at-risk pregnancies, it is important to understand the differences that occur in healthy pregnancies. This study aims to identify whether differences in HRV and PPG morphology between pregnant and non-pregnant women are significant and large enough to discriminate these populations using a machine learning approach. High performance is desired, but the core purpose of the model is to provide insights into features useful for this discrimination. To fulfill the objectives, several data transformation steps were considered to improve the performance of the classifiers. In addition, multiple classifiers were implemented. A pipeline structure was used to identify the best combination of transformations and classifiers. The pipelines input were the features computed from HRV and PPG morphology. A pipeline using logistic regression as classifier was selected as the best model for this application because it represented a good compromise between performance and interpretability. The selected model performed well using ten parameters, ordered by the weights assigned by the logistic regression function. The model provided insights into which features are relevant to this classification and to what extent. These results emphasized the importance of using HRV features that are not commonly included in studies and the added value of combining the study of HRV with PPG morphology features.File | Dimensione | Formato | |
---|---|---|---|
MJAlmario_ExSummary_30_11.pdf
embargo fino al 30/11/2025
Descrizione: Executive summary
Dimensione
830.18 kB
Formato
Adobe PDF
|
830.18 kB | Adobe PDF | Visualizza/Apri |
MJAlmario_Thesis_30_11.pdf
embargo fino al 30/11/2025
Descrizione: Thesis
Dimensione
8.72 MB
Formato
Adobe PDF
|
8.72 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/10589/198045