Cloud-Enabled Time-Series Forecasting for Hospital Readmissions Using Transformer Models and Attention Mechanisms

Abstract

Hospital readmissions are one of the biggest issues in healthcare. They increase the cost and worsen patient outcomes, putting additional strain on already-over-stretched resources. Traditional models for prediction often fail to handle the complexity and high dimensionality of healthcare data. This paper outlines a solution: an approach to the predictive model for hospital readmission using transformer models with attention mechanisms deployed on cloud infrastructure. This approach allows scalability and real-time processing, providing a healthcare professional with more accurate, actionable insights that reduce unnecessary hospitalizations and help in enhancing patient outcomes. The methodology allows the processing of complex time-series data, like vital signs, lab results, and treatment history, thus making the model focus on key features and time periods for improving predictions. The cloud-based infrastructure lets the model scale and adapt dynamically to new data. The experimental results reveal that the proposed model outperforms traditional methods, achieving 88% accuracy, 85% precision, 83% recall, 84% F1-score, and 92% AUC. Future improvements may include integration of genomic data and multi-center deployment.