In the last decades, the growing number of cancer survivors has moved the researchers attention from their cancer treatment to its long-term effects. Adolescents and young adults (AYAs) are a peculiar group of patients with special needs and attentions. Breast cancer (BC) is the most frequent (even though rare) cancer in these patients. BC in young patients is extremely aggressive and so are its treatments. The treatments that offer the promise of BC survival, such as chemotherapy, radiation therapy, and hormonal treatments are often accompanied by potential cardiotoxic effects. These effects can lead to significant cardiovascular disease (CVD) risk, a primary concern given the long life-expectancy of AYA BC survivors. In this context, predicting and explaining the long-term CVD risk associated with BC treatments in AYA is paramount for clinicians to ensure a proactive approach to cardiovascular monitoring and tailored follow-up care. To address this challenge, this project employed a robust methodological framework centered on Bayesian Networks (BNs) and causal inference methods. Bayesian Networks are particularly well-suited for this study due to their ability to represent complex probabilistic relationships and incorporate both clinical knowledge and real-world data (RWD). More in detail, the specific goals of this research were multifaceted: A) Development of a Predictive Causal Model: The primary objective was to design and develop a Bayesian Network that integrates prior medical knowledge and RWD from two cohorts of patients to accurately predict the likelihood of CVDs in AYA breast cancer survivors. In this specific project, handling RWD was challenging due to missing values, selection bias, and class imbalance that required the use of dedicated methodological approaches (such as missingness graphs and Structural Expectation-Maximization algorithm). According to our experimental results, the proposed model performs very good at identifying and characterizing AYA BC patients who are at elevated risk of CVDs, also providing a foundation for informed individualized care strategies. B)Causal Mechanism Exploitation: Beyond prediction, this study sought to uncover the causal pathways through which BC treatments contribute to CVD risk. By leveraging on causal in ference methods, this research aimed to identify the individual role of each treatment, making it evident how all neodjuvant (pre-surgical) treatments are those that mainly influence CVD risk probably because of their interaction with the tumor (not yet surgically removed) and the immune system of the patient. C)Raise awareness about the treatments’ choice: Oncologists base their treatment decision mainly on the cancer prognosis, thus non-oncological long-term effects of the treatments’ choice are often missing in the official treatment guidelines. In our work we showed how the proposed model can be used also to build counterfactual scenarios to enrich standard guidelines for BC treatment with awareness about the CVD risk. This information could be pivotal in case of uncertainty on the treatment choice or for better tailoring the follow-up for high-risk patients. To conclude, by leveraging on BNs properties and causal inference algorithms, in this research project we were able to build a causal model that not only predicts the risk of CVDs in AYA BC survivors but also elucidated the underlying causal mechanisms driving the CVD risk in this peculiar group of patients. Lastly, this work showed a practical use case of how Computer Science methods can be used to addresses a clinical need to ensure that the lives extended by modern cancer treatments are not overshadowed by preventable long-term health issues.
Il numero crescente di sopravvissuti al cancro ha spostato l'attenzione dei ricercatori dal trattamento del tumore allo studio dei suoi effetti a lungo termine. Gli adolescenti e i giovani adulti (AYA) sono un gruppo peculiare di pazienti con esigenze particolari. Il cancro al seno (BC) è il tumore più frequente (seppur raro) nelle AYA. Il BC nelle giovani è aggressivo e richiede, di conseguenza, che lo siano anche i suoi trattamenti. I trattamenti più associati ad un aumento della sopravvivenza, come la chemioterapia, la radioterapia e i trattamenti ormonali, sono spesso accompagnati da potenziali effetti cardiotossici. Questi effetti possono portare a malattie cardiovascolari (CVD), preoccupazione primaria data la lunga aspettativa di vita delle pazienti. Prevedere e spiegare il rischio di CVD a lungo termine associato ai trattamenti per il BC negli AYA è fondamentale per i medici, al fine di garantire un approccio proattivo al monitoraggio e follow-up personalizzati. Per affrontare questa sfida, questo progetto si è focalizzato su reti bayesiane (BN) e metodi di inferenza causale. Le BN sono risultate particolarmente adatte a questo tipo di studio grazie alla loro capacità di rappresentare relazioni probabilistiche complesse e di incorporare sia le conoscenze cliniche che dati del mondo reale (RWD). Più in dettaglio, gli obiettivi specifici di questa ricerca erano molteplici: A) Sviluppo di un modello causale predittivo: L'obiettivo primario di questo progetto quello di sviluppare una BN che integrasse le conoscenze mediche pregresse e i RWD di due coorti di pazienti per prevedere con precisione la probabilità di CVD nelle pazienti AYA sopravvissuti al cancro al seno. In questo progetto specifico, la gestione delle RWD è stata impegnativa a causa dei valori mancanti, dei bias di selezione e della rarità degli eventi. In base ai risultati sperimentali, il modello proposto riesce molto bene ad identificare e caratterizzare le pazienti AYA con BC che sono ad alto rischio di CVD, fornendo anche una base per strategie di cura personalizzate e informate. B) Spiegazione del meccanismo causale: Oltre alla previsione, questo studio ha cercato di indagare i percorsi causali attraverso i quali i trattamenti per il BC contribuiscono al rischio di CVD. Facendo leva sui metodi di inferenza causale, è stato identificato il ruolo individuale di ciascun trattamento, rendendo evidente come tutti i trattamenti neodiuvanti (pre-chirurgici) siano quelli ad influenzare maggiormente il rischio di CVD, probabilmente a causa della loro interazione con il tumore (non ancora rimosso chirurgicamente) e il sistema immunitario della paziente. C) Aumentare la consapevolezza della scelta dei trattamenti: Gli oncologi basano le loro decisioni terapeutiche sulla massimizzazione della prognosi; infatti, gli effetti non oncologici della scelta terapeutica sono spesso assenti nelle linee guida. In questo lavoro abbiamo mostrato come il modello proposto possa essere utilizzato anche per costruire scenari che posso arricchire le linee guida per il trattamento del BC con la consapevolezza del conseguente rischio di CVD. Queste informazioni potrebbero essere fondamentali in caso di incertezza sulla scelta del trattamento o per meglio adattare il follow-up dei pazienti ad alto rischio. In conclusione, sfruttando le proprietà delle BN e gli algoritmi di inferenza causale, siamo stati in grado di costruire un modello causale che non solo predice il rischio di CVD nelle pazienti AYA sopravvissute al BC, ma ha anche chiarito i meccanismi causali sottostanti che determinano il rischio di CVD in questo gruppo di pazienti. Infine, questo lavoro ha mostrato un caso d'uso pratico di come dei metodi informatici possano essere utilizzati per rispondere a un'esigenza clinica, per far sì che, sul lungo termine, le numerose vite allungate dai moderni trattamenti contro il cancro non vengano compromesse da altri problemi di salute prevenibili.
From Real-World Data to Causal Models: A Bayesian Network to study cardiovascular diseases in Adolescents and Young Adults with Breast Cancer
BERNASCONI, ALICE
2025
Abstract
In the last decades, the growing number of cancer survivors has moved the researchers attention from their cancer treatment to its long-term effects. Adolescents and young adults (AYAs) are a peculiar group of patients with special needs and attentions. Breast cancer (BC) is the most frequent (even though rare) cancer in these patients. BC in young patients is extremely aggressive and so are its treatments. The treatments that offer the promise of BC survival, such as chemotherapy, radiation therapy, and hormonal treatments are often accompanied by potential cardiotoxic effects. These effects can lead to significant cardiovascular disease (CVD) risk, a primary concern given the long life-expectancy of AYA BC survivors. In this context, predicting and explaining the long-term CVD risk associated with BC treatments in AYA is paramount for clinicians to ensure a proactive approach to cardiovascular monitoring and tailored follow-up care. To address this challenge, this project employed a robust methodological framework centered on Bayesian Networks (BNs) and causal inference methods. Bayesian Networks are particularly well-suited for this study due to their ability to represent complex probabilistic relationships and incorporate both clinical knowledge and real-world data (RWD). More in detail, the specific goals of this research were multifaceted: A) Development of a Predictive Causal Model: The primary objective was to design and develop a Bayesian Network that integrates prior medical knowledge and RWD from two cohorts of patients to accurately predict the likelihood of CVDs in AYA breast cancer survivors. In this specific project, handling RWD was challenging due to missing values, selection bias, and class imbalance that required the use of dedicated methodological approaches (such as missingness graphs and Structural Expectation-Maximization algorithm). According to our experimental results, the proposed model performs very good at identifying and characterizing AYA BC patients who are at elevated risk of CVDs, also providing a foundation for informed individualized care strategies. B)Causal Mechanism Exploitation: Beyond prediction, this study sought to uncover the causal pathways through which BC treatments contribute to CVD risk. By leveraging on causal in ference methods, this research aimed to identify the individual role of each treatment, making it evident how all neodjuvant (pre-surgical) treatments are those that mainly influence CVD risk probably because of their interaction with the tumor (not yet surgically removed) and the immune system of the patient. C)Raise awareness about the treatments’ choice: Oncologists base their treatment decision mainly on the cancer prognosis, thus non-oncological long-term effects of the treatments’ choice are often missing in the official treatment guidelines. In our work we showed how the proposed model can be used also to build counterfactual scenarios to enrich standard guidelines for BC treatment with awareness about the CVD risk. This information could be pivotal in case of uncertainty on the treatment choice or for better tailoring the follow-up for high-risk patients. To conclude, by leveraging on BNs properties and causal inference algorithms, in this research project we were able to build a causal model that not only predicts the risk of CVDs in AYA BC survivors but also elucidated the underlying causal mechanisms driving the CVD risk in this peculiar group of patients. Lastly, this work showed a practical use case of how Computer Science methods can be used to addresses a clinical need to ensure that the lives extended by modern cancer treatments are not overshadowed by preventable long-term health issues.File | Dimensione | Formato | |
---|---|---|---|
phd_unimib_745907.pdf
accesso aperto
Dimensione
4.9 MB
Formato
Adobe PDF
|
4.9 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/193724
URN:NBN:IT:UNIMIB-193724