The aim of this thesis is to compare two independent samples of multivariate functional data, with correlated components, that differ in term of their variance-covariance operators. One of the major purposes is to recognise the group membership of each observation, performing non-parametric comparison tests as well as supervised classification through generalized regression models. The concept of depth measure is generalized to this kind of data exploiting the role of the variance-covariance operators in weighting the components. Two ways of using the estimated variance-covariance operators are proposed: the first one including both the correlation among different signal components and the covariance between the reciprocal ones, the second one considering the covariance between the same components in the two populations only. It is known that, in general, the realization of an estimated variance-covariance operator is a positive-definite matrix, but an extra-diagonal sub-matrix could be non-positive-definite. For this reason some definitions of distance between positive-definite matrices are generalized to pseudo-distances between non necessarily positive-definite ones. Then the weights of the depth measures are chosen taking into account this type of distance between the estimated variance-covariance operators of two groups. The above-mentioned depth measures are used as predictors in a logistic regression model. In particular, some simulation studies are carried out to validate the robustness of the proposed choice of the weights for the depth measures and a logistic regression model is applied. The same procedure is finally repeated for an application to 8-leads ElectroCardioGraphic signals (ECG) to compare physiological subjects and patients affected by Left Bundle Branch Block.
In questa tesi si confrontano due campioni di dati funzionali multivariati, con componenti correlate, che differiscono in termini di operatori varianza-covarianza e si esegue una classificazione supervisionata tramite modelli di regressione logistica. In quest'ultima i regressori sono le misure di profondità, generalizzate al caso multivariato in modo tale che i pesi utilizzati per calcolarle tengano in considerazione la distanza tra operatori varianza-covarianza di gruppi diversi. Sono proposti due metodi di utilizzo di tali operatori: il primo include sia la correlazione tra componenti diverse del segnale sia la covarianza tra quelle reciproche, il secondo considera solo la covarianza tra le stesse componenti nelle due popolazioni. In generale, la realizzazione di un operatore varianza-covarianza campionaria è una matrice definita positiva, ma le sue sottomatrici extradiagonali non è detto che lo siano. A questo proposito si generalizzano le definizioni delle distanze ben poste per matrici definite positive al caso di pseudo-distanze per matrici che possono non esserlo. Si esegue, in primo luogo, una simulazione di dati bivariati finalizzata a mostrare la robustezza sia del metodo di calcolo del peso proposto sia della classificazione risultante dalla regressione logistica. Si ripete, in secondo luogo, lo stesso procedimento in un caso applicativo reale, nel quale i dati funzionali a disposizione siano elettrocardiogrammi a 8 derivazioni di soggetti sani ed affetti da LBBB (Left Bundle Branch Block).
Uso delle misure di profondità per dati funzionali multivariati nella previsione di patologie : un'applicazione ai segnali elettrocardiografici
BIASI, RACHELE
2012/2013
Abstract
The aim of this thesis is to compare two independent samples of multivariate functional data, with correlated components, that differ in term of their variance-covariance operators. One of the major purposes is to recognise the group membership of each observation, performing non-parametric comparison tests as well as supervised classification through generalized regression models. The concept of depth measure is generalized to this kind of data exploiting the role of the variance-covariance operators in weighting the components. Two ways of using the estimated variance-covariance operators are proposed: the first one including both the correlation among different signal components and the covariance between the reciprocal ones, the second one considering the covariance between the same components in the two populations only. It is known that, in general, the realization of an estimated variance-covariance operator is a positive-definite matrix, but an extra-diagonal sub-matrix could be non-positive-definite. For this reason some definitions of distance between positive-definite matrices are generalized to pseudo-distances between non necessarily positive-definite ones. Then the weights of the depth measures are chosen taking into account this type of distance between the estimated variance-covariance operators of two groups. The above-mentioned depth measures are used as predictors in a logistic regression model. In particular, some simulation studies are carried out to validate the robustness of the proposed choice of the weights for the depth measures and a logistic regression model is applied. The same procedure is finally repeated for an application to 8-leads ElectroCardioGraphic signals (ECG) to compare physiological subjects and patients affected by Left Bundle Branch Block.| File | Dimensione | Formato | |
|---|---|---|---|
|
2013_12_Biasi.pdf
accessibile in internet per tutti
Dimensione
15.89 MB
Formato
Adobe PDF
|
15.89 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/10589/88589