Tesi etd-09282026-100759 |
Link copiato negli appunti
Tipo di tesi
Tesi di laurea magistrale LM6
Autore
CRISTOFANI, DAVID
URN
etd-09282026-100759
Titolo
AI-Driven TAVR vs SAVR Selection: Towards Personalized Aortic Valve Replacement
Dipartimento
RICERCA TRASLAZIONALE E DELLE NUOVE TECNOLOGIE IN MEDICINA E CHIRURGIA
Corso di studi
MEDICINA E CHIRURGIA
Relatori
.
relatore Colli, Andrea
correlatore Besola, Laura
correlatore Besola, Laura
Parole chiave
- IA
- SAVR
- TAVR
Data inizio appello
20/10/2026
Consultabilità
Non consultabile
Data di rilascio
20/10/2029
Riassunto (Inglese)
Methods: Aggregated data from 67 clinical studies were included, totaling 65,742 real-world patients. We reconstructed microdata populations ex novo using parallel Bayesian chain and iterative proportional fitting algorithms constrained by empirical comorbidity priors. We used five independent stratified cohorts (n=10,000 each) to train Classification and Regression Tree (CART) models and gradient-boosted ensembles. We interpreted algorithmic choices using SHapley Additive exPlanations (SHAP) and calibrated them against multivariable logistic regression models.
Results: Generative reconstruction achieved high marginal fidelity (global mean absolute error 0.0335). Tree-based classifiers demonstrated stable cross-validation accuracy across synthetic folds (75.1%). Age was the primary root partition, resulting in SAVR for patients <75 years and TAVR for those >82 years. In the 75-82-year grey zone, comorbidities influenced the selection: renal impairment strongly shifted selection toward TAVR, while concomitant coronary disease and diabetes favoured SAVR. Long-term attrition independently predicted 10-year mortality (OR: 92.14, p<0.001) and structural reintervention (OR: 1.47, p=0.017) in favour of SAVR.
Conclusion: Our model identified empirical age thresholds favouring TAVR in patients aged >82 years and SAVR in those aged <75 years. Renal function and macrovascular disease strongly influenced procedural choice within the 75-82 years grey zone. XAI trained on generative synthetic populations establishes a transparent, reproducible framework for individualized Heart Team shared decision-making; however, prospective external validation on granular real-world registries remains essential to confirm clinical translation.
Results: Generative reconstruction achieved high marginal fidelity (global mean absolute error 0.0335). Tree-based classifiers demonstrated stable cross-validation accuracy across synthetic folds (75.1%). Age was the primary root partition, resulting in SAVR for patients <75 years and TAVR for those >82 years. In the 75-82-year grey zone, comorbidities influenced the selection: renal impairment strongly shifted selection toward TAVR, while concomitant coronary disease and diabetes favoured SAVR. Long-term attrition independently predicted 10-year mortality (OR: 92.14, p<0.001) and structural reintervention (OR: 1.47, p=0.017) in favour of SAVR.
Conclusion: Our model identified empirical age thresholds favouring TAVR in patients aged >82 years and SAVR in those aged <75 years. Renal function and macrovascular disease strongly influenced procedural choice within the 75-82 years grey zone. XAI trained on generative synthetic populations establishes a transparent, reproducible framework for individualized Heart Team shared decision-making; however, prospective external validation on granular real-world registries remains essential to confirm clinical translation.
Riassunto (Italiano)
File
| Nome file | Dimensione |
|---|---|
La tesi non è consultabile. |
|