In this thesis we analyse the impact of AntiRetroviral Therapies (ARTs) drugs and other clinical measurements on the time to a CardioVascular Diseases (CVDs) event in HIV patients through the application of survival methods to a real dataset. We consider two different approaches: the classical Proportional Hazard (PH) Cox model and a neural network based method, the DeepHit. First, we analyse and compare the two methods fitted on clinical and therapies data measured at the baseline, i.e., at the beginning of the ART. Then, we move to a time-dependent setting, considering the whole follow-up of patients, and we analyse and compare extensions of the two methods: the time-dependent Cox PH model and the Dynamic DeepHit. All methods are compared in terms of interpretability and predictive performances. The compared models have similar performances and are able to reach high values of Concordance-index (0.77). The neural network method is more flexible than the Cox model, it relaxes the PH assumption and allows to capture non-linear and time-varying relationships between the covariates and the target variable. On the other side, it has two weaknesses: the computational cost, that becomes prohibitive with time-dependent data, and the difficulty of interpretation. This last problem is turned into a point of strength with the use of the Shapley Additive Explanation that enabled to interpret and visualise the interaction between covariates and the time to CVD events. These models provide very interesting results: short time of exposure to ART inhibitor drugs increases the risk of CVD events in 15 years, while long time of exposure to these drugs is a strong protective factor.
In questa tesi, analizziamo l’impatto che specifici medicinali usati nelle Terapie AntiRetrovirali (ART) e le caratteristiche cliniche e personali di pazienti affetti da HIV hanno sul rischio di un evento cardiovascolare, applicando metodi di analisi di sopravvivenza su dati reali. Consideriamo due approcci differenti: il classico modello Cox Proportional Hazard (PH) e il DeepHit, un metodo basato su reti neurali. Prima, analizziamo e confrontiamo i due metodi considerando i dati clinici e della terapia alla baseline, i.e. i dati misurati all’inizio della ART. In seguito, includiamo dati dipendenti dal tempo, misurati sullo storico delle visite dei pazienti, e applichiamo e confrontiamo due altri metodi: il modello di Cox tempo-dipendente e il Dynamic DeepHit. I modelli vengono valutati e confrontati in termini di interpretabilità e capacità predittiva. I due approcci mostrano performance predittive simili e raggiungono valori alti di C-index (0.77). Il metodo basato su reti neurali permette di rilassare l’ipotesi di PH e di linearità degli effetti delle covariate, risultando più flessibile del modello di Cox. Dall’altro lato, presenta due debolezze: il costo computazionale, che diventa proibitivo con i dati tempo dipendenti, e la difficoltà di interpretazione dei risultati. L’applicazione delle tecniche di interpretazione degli Shapley values, che permettono di interpretare e visualizzare l’interazione tra due variabili e il tempo all’evento cardiovascolare, trasforma la seconda debolezza in un punto di forza. I modelli applicati producono interessanti risultati: una breve esposizione ai medicinali delle ART aumenta di poco il rischio di un evento cardiovascolare a 15 anni, mentre una lunga esposizione diminuisce fortemente questo rischio.
A neural network approach to survival analysis with time-dependent covariates for modelling time to cardiovascular diseases in HIV patients
LURANI CERNUSCHI, AGOSTINO
2020/2021
Abstract
In this thesis we analyse the impact of AntiRetroviral Therapies (ARTs) drugs and other clinical measurements on the time to a CardioVascular Diseases (CVDs) event in HIV patients through the application of survival methods to a real dataset. We consider two different approaches: the classical Proportional Hazard (PH) Cox model and a neural network based method, the DeepHit. First, we analyse and compare the two methods fitted on clinical and therapies data measured at the baseline, i.e., at the beginning of the ART. Then, we move to a time-dependent setting, considering the whole follow-up of patients, and we analyse and compare extensions of the two methods: the time-dependent Cox PH model and the Dynamic DeepHit. All methods are compared in terms of interpretability and predictive performances. The compared models have similar performances and are able to reach high values of Concordance-index (0.77). The neural network method is more flexible than the Cox model, it relaxes the PH assumption and allows to capture non-linear and time-varying relationships between the covariates and the target variable. On the other side, it has two weaknesses: the computational cost, that becomes prohibitive with time-dependent data, and the difficulty of interpretation. This last problem is turned into a point of strength with the use of the Shapley Additive Explanation that enabled to interpret and visualise the interaction between covariates and the time to CVD events. These models provide very interesting results: short time of exposure to ART inhibitor drugs increases the risk of CVD events in 15 years, while long time of exposure to these drugs is a strong protective factor.File | Dimensione | Formato | |
---|---|---|---|
2021_12_Lurani_Cernuschi.pdf
Open Access dal 20/11/2022
Dimensione
3.23 MB
Formato
Adobe PDF
|
3.23 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/10589/182180