Sparse Neural Networks via Bi-level ℓ0 Optimization for Prostate Cancer Prognosis

This study proposes a bi-level optimization framework integrating ℓ0 regularization into deep neural networks for joint model compression and embedded feature selection. The approach employs Hard Concrete distributions to approximate the nonconvex ℓ0 norm, enabling simultaneous learning of binary masks for neurons and input features within a differentiable setting. Applied to a cohort of prostate cancer patients, the framework effectively identifies key clinical and molecular predictors while achieving substantial sparsity. Experimental results show 70% neuron pruning and 60% feature reduction with 90% validation accuracy. The method enhances both computational efficiency and clinical interpretability, providing a scalable foundation for decision-support applications in oncology.