Survival analysis is central to medical research and clinical decision-making, where estimating time-to-event outcomes is essential for prognosis, risk stratification, and treatment planning. Medical survival datasets are often high-dimensional, censored, and heterogeneous, creating challenges for both predictive accuracy and interpretability. While deep learning–based survival models achieve strong predictive performance, their black-box nature limits clinical applicability, whereas clustering-based approaches provide interpretability at the cost of accuracy. This thesis introduces CONVERSE (CONtrastive Variational Ensemble for Risk Stratification and Estimation), a clustering-aware deep survival framework that integrates representation learning, latent-space clustering, and survival prediction. CONVERSE supports deterministic, Siamese, and variational autoencoders, and incorporates contrastive and self-paced learning to encourage meaningful latent structures, while allowing multiple clustering strategies and optional cluster-specific survival heads. The framework is evaluated on multiple medical survival datasets. Results show that CONVERSE achieves competitive performance, often outperforming existing clustering-based methods and matching classical survival models. Although purely machine learning–based approaches may achieve higher accuracy in some settings, qualitative analysis demonstrates that CONVERSE learns clinically meaningful subgroups, highlighting the trade-off between interpretability and peak performance and motivating flexible, modular model design.
L’analisi di sopravvivenza è centrale nella ricerca medica e nei processi decisionali clinici, dove le stime temporali sono essenziali per la prognosi, l'identificazione dei rischi e la pianificazione del trattamento. I dataset medici sono spesso ad alta dimensionalità, incompleti ed eterogenei, creando sfide sia per l’accuratezza predittiva sia per l’interpretabilità dei risultati ottenuti. Sebbene i modelli di sopravvivenza basati sul deep learning raggiungano elevate prestazioni predittive, la loro natura black box ne limita l’applicabilità clinica, mentre gli approcci basati sul clustering offrono interpretabilità a scapito dell’accuratezza. Questa tesi introduce CONVERSE (CONtrastive Variational Ensemble for Risk Stratification and Estimation), un modello di sopravvivenza basato sul clustering che integra apprendimento delle stratificazioni di pazienti, clustering nello spazio latente e predizione della sopravvivenza. CONVERSE supporta autoencoder deterministici, siamesi e variazionali, e incorpora contrastive e self-paced learning per favorire strutture latenti significative, consentendo al contempo diverse strategie di clustering e stime di sopravvivenza specifiche per cluster. Il modello è valutato su diversi dataset medici. I risultati mostrano che CONVERSE raggiunge prestazioni competitive, spesso superando i metodi basati sul clustering esistenti e eguagliando i modelli classici di sopravvivenza. Sebbene approcci puramente basati sul machine learning possano ottenere una maggiore accuratezza in alcuni contesti, l’analisi qualitativa dimostra che CONVERSE apprende sottogruppi clinicamente significativi, favorendo il compromesso tra interpretabilità e prestazioni.
CONVERSE: discovering patient groups in survival data via deep latent clustering
ERBIL, PINAR
2025/2026
Abstract
Survival analysis is central to medical research and clinical decision-making, where estimating time-to-event outcomes is essential for prognosis, risk stratification, and treatment planning. Medical survival datasets are often high-dimensional, censored, and heterogeneous, creating challenges for both predictive accuracy and interpretability. While deep learning–based survival models achieve strong predictive performance, their black-box nature limits clinical applicability, whereas clustering-based approaches provide interpretability at the cost of accuracy. This thesis introduces CONVERSE (CONtrastive Variational Ensemble for Risk Stratification and Estimation), a clustering-aware deep survival framework that integrates representation learning, latent-space clustering, and survival prediction. CONVERSE supports deterministic, Siamese, and variational autoencoders, and incorporates contrastive and self-paced learning to encourage meaningful latent structures, while allowing multiple clustering strategies and optional cluster-specific survival heads. The framework is evaluated on multiple medical survival datasets. Results show that CONVERSE achieves competitive performance, often outperforming existing clustering-based methods and matching classical survival models. Although purely machine learning–based approaches may achieve higher accuracy in some settings, qualitative analysis demonstrates that CONVERSE learns clinically meaningful subgroups, highlighting the trade-off between interpretability and peak performance and motivating flexible, modular model design.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Erbil_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Text of the executive summary
Dimensione
1.55 MB
Formato
Adobe PDF
|
1.55 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Erbil_Thesis.pdf
accessibile in internet per tutti
Descrizione: Text of the thesis
Dimensione
4.76 MB
Formato
Adobe PDF
|
4.76 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/10589/252058