A Multimodal Rolling-Window Framework for ICU Transfer Prediction in Hospitalized Patients with Comorbid Hypertension

Hospitalized patients with comorbid systemic hypertension represent a high-risk population prone to clinical deterioration. Early detection is crucial for improving prognosis and optimizing the allocation of intensive care resources. However, dynamic risk prediction tools for this high-risk cohort are still limited. Therefore, this study developed a new universal risk prediction framework that matches the decision-making rhythm of daily ward rounds. This framework adopts a multimodal rolling-window design, with window set at 24-hour intervals spanning from 24 to 264 hours after admission. First, it integrates unstructured clinical text features (e.g., chief complaints and past medical history, processed by large language models) with multimodal structured features, encompassing static elements (demographics and administrative metadata) and irregularly sampled time-series variables (e.g., laboratory and microbiological data), modeled by traditional machine learning. Then, independent samples are constructed in each anchor window and anchor-specific training is carried out to achieve dynamic identification and prospective prediction of patients at high risk of ICU transfer within the next 72 hours. In the large-scale retrospective cohort of MIMIC IV database, the framework demonstrated excellent dynamic prediction performance, with an AUPRC of \(\:0.4650\pm\:0.0898\), an AUROC of \(\:0.9077\pm\:0.0362\) and a sensitivity of \(\:0.8423\:\pm\:\:0.0164\)The experiment confirmed that multimodal data fusion can significantly improve predictive performance. Furthermore, SHAP-based interpretability analysis revealed that the model captured a sequential decompensation pattern, from initial inflammatory activation to multisystem dysfunction and eventual systemic metabolic imbalance, enhancing the clinical credibility of the model. In summary, this study provides an extensible and interpretable dynamic risk prediction framework for hospitalized hypertensive patients with multiple comorbidities. This framework is expected to support clinical proactive decision-making and provide data-driven new strategies for optimizing critical care resource allocation.