Enhancing Result Interpretability of Neural Architecture Search-Assisted Medical AI via Large Language Model

NAS can find better solutions by exploring the vast search space for possible architectures and is widely used in medical AI. However, due to its black-box nature, explainability is important for NAS-assisted medical AI applications. In this paper, we design and implement a risk prediction model for amyloidosis (a rare hematological disease) using an evolutionary neural architecture search (ENAS) to automatically find the best model architecture. To improve the professional acceptance of this prediction model among clinicians, we introduce the concept of result interpretability. Meanwhile, we propose a framework for medical AI systems that can interpret results, utilizing a large language model. The realization of the system provides a more feasible way to apply AI-assisted medical diagnosis on a large scale.