Gamt-Care: A Graph Attention Multimodal Transformer Framework for Crm-Augmented Predictive Healthcare Analytics

Abstract

The explosion of digital healthcare services has resulted in massive amounts of patient data being collected by hospitals through their CRM systems, including appointment records, follow-ups, and communications with patients about their treatment. Unfortunately, the vast majority of predictive healthcare models today primarily utilize EHR data and largely ignore the behavioral insights that are present in CRM interaction data. To overcome this limitation, this study developed a new intelligent framework for predictive healthcare analytics and AI-enhanced clinical workflow management called GAMT-CARE (Graph Attention Multimodal Transformer for CRM-Augmented Risk Estimation). This framework introduces a new type of patient interaction graph (PIG) that models the various relationships between patients, providers, treatments, and post-treatment communications as they relate to the CRM systems used by hospitals. In order to effectively analyze the heterogeneous data associated with the healthcare ecosystem, a hybrid deep learning architecture known as the GAMTNet (Graph Attention Multimodal Transformer Network) was developed. This model utilizes medical imaging features computed using EfficientNetV2, temporal patient histories modelled using bi-directional long short term memory (Bi-LSTM), and CRM interaction patterns learned through graph neural networks (GNNs) to create a single multimodal representation of each patient's clinical, behavioral, and temporal characteristics that can be fused together using a transformer-based attention mechanism to model the complex relationships among these three different types of features. Additionally, a hybrid swarm intelligence optimization strategy based on the Adaptive Predator Swarm Optimization (APSO), a hybrid bio-inspired optimization algorithm that combines the aggressive swarm behavior of red piranhas (RPO) with the cooperative hunting strategy of grey wolves (GWO) for efficient feature selection and hyperparameter tuning, improves model performance. The suggested approach supports intelligent clinical workflow optimization and individualized healthcare intervention by enabling early prediction of patient risk levels, disease development, and hospital readmission probability.