Abstract
Abnormal serum sodium levels are a common and severe complication in stroke patients, significantly increasing mortality risk and prolonging ICU stays. Accurate real-time prediction of serum sodium fluctuations is crucial for optimizing clinical interventions. However, existing predictive models face limitations in handling complex dynamic features and long time series data, making them less effective in guiding individualized treatment. To address this challenge, this study developed a deep learning model based on a multi-head attention mechanism to enable real-time prediction of serum sodium concentrations and provide personalized intervention recommendations for ICU stroke patients. This study utilized publicly available MIMIC-III (n = 2346) and MIMIC-IV (n = 896) datasets, extracting time series data from 10 key clinical indicators closely associated with serum sodium levels. To address the complexity of long time series data, a moving sliding window sub-sampling segmentation method was employed, effectively transforming extensive sequences into more manageable inputs while preserving critical temporal dependencies. By leveraging advanced mathematical modeling, meaningful insights were extracted from sparse and irregular time series data. The resulting time-feature fusion multi-head attention (TFF-MHA) model underwent rigorous validation using public datasets and demonstrated superior performance in predicting both serum sodium values and corresponding intervention measures compared to existing models. This study contributes to the field of healthcare informatics by introducing an innovative, data-driven approach for dynamic serum sodium prediction and intervention recommendation, providing a valuable clinical decision-support tool for optimizing sodium management strategies in critically ill stroke patients.