A deep learning approach to predicting hospitalized patients' SEIRD states using multimodal spatiotemporal data

Abstract

BACKGROUND AND OBJECTIVE: The increasing prevalence of hospital-acquired infections (HAIs) due to antimicrobial resistance presents a formidable challenge to patient outcomes and resource allocation. Existing prediction models often operate at a population level, failing to provide the granular, individual patient-specific risk assessments crucial for targeted interventions. Our study addresses this gap by developing and evaluating multimodal deep learning models designed to leverage rich spatial and temporal hospital data for precise, individual-level HAI risk prediction. METHODS: We engineered three distinct deep learning architectures, each employing a different strategy for spatiotemporal data integration. These included: 1) an integrated approach using Heterogeneous Graph Convolution Long Short-Term Memory for unified learning; 2) a decoupled model combining separate Long Short-Term Memory and Diffusion Convolutional Recurrent Neural Network components; and crucially, 3) a novel hybrid model. This hybrid architecture is designed to first allow specialized components to learn distinct representations from spatial and temporal data independently, followed by a joint fine-tuning phase that intelligently fuses these pre-trained representations. All models were rigorously evaluated using a peer-reviewed synthetic hospital simulation dataset that meticulously captures real-world patient movement and infection dynamics. We employed stratified 10-fold cross-validation with accuracy and F1 score metrics. RESULTS: The hybrid model consistently surpassed both the integrated and decoupled paradigms across all experimental conditions. Over a 7-day prediction window, it achieved peak accuracy (78.96 %) and F1 score (0.75), substantially outperforming the decoupled (71.46 % accuracy, 0.72 F1) and integrated (57.73 % accuracy, 0.42 F1) models. Furthermore, the hybrid model demonstrated enhanced generalization and robustness, maintaining strong performance across varying hospital scales and prediction horizons. CONCLUSIONS: Our findings underscore the efficacy of a hybrid deep learning strategy that skillfully combines specialized learning of spatial and temporal hospital data with a subsequent joint fine-tuning process. This innovative approach yields superior predictive capabilities for individual patient HAI risk, even when evaluated on a complex, real-world representative synthetic dataset. The demonstrated performance and methodological insights suggest significant potential for real-time clinical decision support and optimization of infection control measures. Its inherent adaptability makes it a promising foundation for deployment in diverse healthcare settings, with future work focused on validation with real-world clinical data.