Explainable Clinical AI Decision-Making for Myeloproliferative Neoplasms: Integration of Consumer-Grade Wearable Monitors for Real-Time Risk Assessment and Treatment Optimization

Prediction of risks and constant monitoring is essential in treatment of Myeloproliferative Neoplasms (MPNs). The design of this paper aims at providing an Explainable Artificial Intelligence-based framework, integrated with consumer-level wearable devices can enable accurate, low-latency, and clinically trusted real-time detection of thrombotic risk and treatment optimization. The system uses multimodal signal acquisition that includes the photoplethysmography (PPG), heart rate variability (HRV), accelerations, and clinical biomarkers. A hybrid CNN-Transformer backbone is employed for temporal and contextual feature fusion, with Grad-CAM and SHAP used to generate visual and feature-level explanations supporting clinical interpretability. The mathematical model formalizes the preprocessing, fusion, interpretability, and multi-objective optimization of the accuracy, latency, safety, and clinician trust. Training and validation were done using synthetic and clinical wearable datasets. Experimental evaluation demonstrates a risk prediction accuracy of 92.8%, an event detection AUC of 0.94, an inference latency of 1.8 s, and clinician trust alignment of 89%, outperforming conventional LSTM- and Transformer-based baselines. The hybrid architecture of edges and clouds proposed provides guarantee of low-latency inference, secure synchronization, and regulatory compliance. All in all, the research establishes interpretable, efficient, and clinically dependable wearable AI system to manage the risk of MPN and personalized treatment support.