A Multi-Source Time Series Hybrid Prediction Framework for Early Warning of Dialysis Access Stenosis

Dialysis access stenosis represents a prevalent and clinically significant complication among maintenance hemodialysis patients, posing a substantial threat to treatment continuity and prognosis. Current monitoring ways are primarily depending on physical examinations and single-point imaging information, which are unable to track the dynamic risk. In this situation, we propose the Dual-phase Access Prediction Network (DAP-Net), a two-stage predictive framework based on multivariate time series modeling. By combining dialysis session records with quarterly blood biomarker data, DAP-Net sequentially estimates the risk of intimal hyperplasia and subsequent thrombosis events. Input features are organized into fixed-length time series segments using a sliding window strategy and enhanced by multi-dimensional time encoding. Experimental evaluation results are shown that appropriate tuning of the sliding window length substantially improves performance: five-fold cross-validation yields maximum accuracies of $\mathbf{8 4. 3 0 \%}$ for intimal hyperplasia prediction and 86.10% for thrombosis prediction. SHAP-based interpretability analysis corroborates the clinical relevance of key features-such as age, blood flow rate, and inflammatory markers-thereby enhancing the transparency and credibility of the model. Overall, DAP-Net holds significant promise for early warning and dynamic management of dialysis access stenosis.