Heart failure (HF) is still a major cause of morbidity and death worldwide, hence prompt intervention requires early
detection. In this work, we suggest a unique hybrid deep learning architecture that uses clinical data to predict the early development of heart failure by combining supervised and unsupervised learning. Our design uses a deep neural network (DNN) for risk categorization after a deep autoencoder for denoised feature
extraction.
Unlike prior works that focus solely on predictive accuracy,we incorporate formal logic-based safety constraints using the Marabou verification framework, enabling our model to operate with mathematical safety guarantees an essential step for real world deployment in clinical settings. Specifically, we verify two
critical properties: output stability (resistance to small input perturbations) and bounded input robustness (reliable behavior
within clinically valid input ranges).
The model achieves a classification accuracy of 99.45%,validated through stratified 10-fold cross-validation and repeated trials to ensure statistical robustness. Our results demonstrate that integrating formal methods with deep learning significantly improves both performance and safety, highlighting the framework’s potential for deployment in trustworthy clinical AI systems.
Index Terms—Safety, Formally, Verified, Heart, Failure, DNN

Auteurs : Sarra Abidi , Imen Chebbi, Leila Ben Ayed

Téléchargement : PDF