Chimeric Antigen Receptor T-cell (CAR-T) therapy has revolutionized the treatment of relapsed or refractory B-Cell Lymphoma, but durable remission is not observed in all patients. Early identification of subjects at high risk of disease progression is crucial. This study evaluates static and dynamic Machine Learning (ML) and Deep Lerning (DL) architectures to predict Progression-Free Survival (PFS) in a real-world multicenter cohort of 1,465 patients enrolled in the CART-SIE study. I utilized clinical and laboratory variables collected at leuka pheresis , infusion, and subsequent standardized follow-ups (between 30 and 180 days post-infusion). For static prediction, Cox Proportional Hazards (CPH) and MLmodelswere compared with deep neural networks (DeepSurv). Although Deep Surv achieved superior internal discriminative performance, the standard CPH model maintained a higher C-index on the independent external validation cohort (0.73 versus 0.69), demonstrating greater generalizability. To leverage longitudi nal data, dynamic predictive frameworks were developed, comparing Joint Mod els, Dynamic-DeepHit and superlandmarking strategies based on Recurrent Neural Networks (RNN). Results demonstrate that encoding the patient’s entire trajec tory through a hybrid RNN-CPH architecture offers the best predictive capabilities. Comparisons among all models were performed using time-variant AUC (tvAUC) at leukapheresis on both test and external validation sets. In these evaluations, the RNN-CPH architecture achieved achieved superior results, likely because its recurrent encoder successfully learned from the temporal evolution of patient tra jectories In conclusion, while the standard CPH confirms itself as a solid model for static predictions, the integration of sequential clinical measurements through hybrid recurrent architectures within a super-landmarking framework offers a sub stantial advantage for continuous, dynamic risk stratification in CAR-T therapy.
La terapia con cellule T con recettore chimerico dell’antigene (CAR-T) ha rivoluzionato il trattamento del linfoma a cellule B recidivato o refrattario, ma non tutti i pazienti raggiungono una remissione duratura. L’identificazione precoce dei soggetti ad alto rischio di progressione della malattia è cruciale. Questo studio valuta architetture statiche e dinamiche di Machine Learning (ML) e Deep Learning (DL) per predire la Progression-Free Survival (PFS) in una coorte multicentrica di 1.465 pazienti arruolati nello studio CART-SIE. Abbiamo utilizzato variabili cliniche e di laboratorio raccolte alla leucaferesi, all’infusione e nei successivi follow up (tra 30 e 180 giorni dopo l’infusione). Per la predizione statica, i modelli di Cox Proportional Hazards (CPH) e i modelli di ML sono stati confrontati con reti neurali profonde (DeepSurv). Sebbene Deep Surv abbia ottenuto una performance discriminativa interna superiore, il modello CPH standard ha mantenuto un C-index più elevato nella coorte indipendente di validazione esterna (0,73 vs 0,69), dimostrando una maggiore generalizzabilità. Per sfruttare i dati longitudinali, sono stati sviluppati framework predittivi dinam ici, confrontando Joint Models, Dynamic-DeepHit e strategie di super-landmarking basate su Recurrent Neural Networks (RNN). I risultati dimostrano che codificare l’intera traiettoria del paziente attraverso un’architettura ibrida RNN-CPH offre le migliori capacità predittive. I confronti tra tutti i modelli sono stati effettuati uti lizzando la AUC tempo-variante (tvAUC) alla leucaferesi sia sul set di test sia sul set di validazione esterna. In queste valutazioni, l’architettura RNN-CPH ha ottenuto risultati superiori, probabilmente perché il suo en coder ricorrente ha appreso con successo dall’evoluzione temporale delle traiettorie dei pazienti. In conclusione, mentre il CPH standard si conferma un modello solido per le predizioni statiche, l’integrazione di misurazioni cliniche sequenziali tramite architetture ricorrenti ibride all’interno di un framework di super-landmarking offre un vantaggio sostanziale per una stratificazione del rischio continua e dinamica nella terapia CAR-T.
Longitudinal machine learning models for CAR-T therapy outcome in large b-cell lymphoma
Simoni, Marco
2024/2025
Abstract
Chimeric Antigen Receptor T-cell (CAR-T) therapy has revolutionized the treatment of relapsed or refractory B-Cell Lymphoma, but durable remission is not observed in all patients. Early identification of subjects at high risk of disease progression is crucial. This study evaluates static and dynamic Machine Learning (ML) and Deep Lerning (DL) architectures to predict Progression-Free Survival (PFS) in a real-world multicenter cohort of 1,465 patients enrolled in the CART-SIE study. I utilized clinical and laboratory variables collected at leuka pheresis , infusion, and subsequent standardized follow-ups (between 30 and 180 days post-infusion). For static prediction, Cox Proportional Hazards (CPH) and MLmodelswere compared with deep neural networks (DeepSurv). Although Deep Surv achieved superior internal discriminative performance, the standard CPH model maintained a higher C-index on the independent external validation cohort (0.73 versus 0.69), demonstrating greater generalizability. To leverage longitudi nal data, dynamic predictive frameworks were developed, comparing Joint Mod els, Dynamic-DeepHit and superlandmarking strategies based on Recurrent Neural Networks (RNN). Results demonstrate that encoding the patient’s entire trajec tory through a hybrid RNN-CPH architecture offers the best predictive capabilities. Comparisons among all models were performed using time-variant AUC (tvAUC) at leukapheresis on both test and external validation sets. In these evaluations, the RNN-CPH architecture achieved achieved superior results, likely because its recurrent encoder successfully learned from the temporal evolution of patient tra jectories In conclusion, while the standard CPH confirms itself as a solid model for static predictions, the integration of sequential clinical measurements through hybrid recurrent architectures within a super-landmarking framework offers a sub stantial advantage for continuous, dynamic risk stratification in CAR-T therapy.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Simoni_ExecutiveSummary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
693.79 kB
Formato
Adobe PDF
|
693.79 kB | Adobe PDF | Visualizza/Apri |
|
2026_03_Simoni_Tesi.pdf
accessibile in internet per tutti
Descrizione: Testo Tesi
Dimensione
2.58 MB
Formato
Adobe PDF
|
2.58 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/10589/252878