MSKT: multimodal data fusion for improved nursing management in hemorrhagic stroke.

Abstract

Background: The study aims to address the challenges of nursing decision-making and the optimization of personalized nursing plans in the management of hemorrhagic stroke. Due to the rapid progression and high complexity of hemorrhagic stroke, traditional nursing methods struggle to cope with the challenges posed by its high incidence and high disability rate.

Methods: To address this, we propose an innovative approach based on multimodal data fusion and a non-stationary Gaussian process model. Utilizing multidimensional data from the MIMIC-IV database (including patient medical history, nursing records, laboratory test results, etc.), we developed a hybrid predictive model with a multiscale kernel transformer non-stationary Gaussian process (MSKT-NSGP) architecture to handle non-stationary time-series data and capture the dynamic changes in a patient's condition.

Results: The proposed MSKT-NSGP model outperformed traditional algorithms in prediction accuracy, computational efficiency, and uncertainty handling. For hematoma expansion prediction, it achieved 85.5% accuracy, an area under the curve (AUC) of 0.87, and reduced mean squared error (MSE) by 18% compared to the sparse variational Gaussian process (SVGP). With an inference speed of 55 milliseconds per sample, it supports real-time predictions. The model maintained a confidence interval coverage near 95% with narrower widths, indicating precise uncertainty estimation. These results highlight its potential to enhance nursing decision-making, optimize personalized plans, and improve patient outcomes.