ThyroDiff-X: A Diffusion-Driven Explainable AI Framework for Early Thyroid Cancer Detection Using Smart Health Biomarkers

Abstract

Smart health devices are playing an increasingly vital role in early diagnosis and personalized care, especially for complex conditions like thyroid cancer. Yet, making sense of subtle, time-dependent patterns in patient data remains challenging— particularly when diagnostic accuracy, transparency, and real-time responsiveness all matter. To tackle this, we introduce the Attention-driven Quantum-Explainable Unified Block with integrated Diffusion Modeling (ThyroDiff-X). ThyroDiff-X brings together three complementary components to model evolving clinical trends. This integration enables the model to capture longitudinal biomarker evolution and suppress noise through diffusion-based temporal reconstruction, explaining its strong performance gains over static and single-visit diagnostic approaches. Specifically, quantum-inspired sinusoidal encoding enriches feature representation, a context-aware diffusion module (Enhanced-DDIM) reconstructs temporal biomarker trajectories using visit-level memory and SHAP-informed clinical context, and multi-head attention adaptively focuses on the most relevant visits. A built-in SHAP layer adds interpretability by highlighting the key features behind each prediction. To optimize performance, the model leverages MSAMO-VCS—a hybrid tuning approach that combines wide-ranging parameter exploration with precise local refinement. The model was rigorously tested across four diverse datasets—SmartThyro-300, UCI Thyroid, Sick-Euthyroid, and THYROID-MVTS—where it consistently outperformed leading baselines, achieving up to 99.5% accuracy, a log loss as low as 0.04, and the highest Quantum Explainability Confidence Rate (QXCR). Its successful deployment on a Jetson Nano board confirmed its real-time readiness and efficiency under resource-constrained settings.

Ähnliche Arbeiten