AI-Driven Analysis of C-CDA Discharge Summaries for Predicting Hospital Readmission and Optimizing Transitions of Care Through FHIR-Enabled Clinical Interoperability

Abstract

Hospital readmission following discharge remains a persistent challenge because clinically relevant risk indicators are distributed across narrative discharge summaries, medication histories, diagnostic findings, follow-up instructions, social factors, and fragmented health-information systems. This study proposes an artificial-intelligence framework for predicting hospital readmission and strengthening transitions of care through integrated analysis of Consolidated Clinical Document Architecture (C-CDA) discharge summaries and Fast Healthcare Interoperability Resources (FHIR)-enabled clinical data. A novel algorithm, termed C2FHIR-ReadmitNet, is developed to combine transformer-based clinical language representation, structured FHIR resource embeddings, temporal transition modelling, and cross-modal attention for patient-level readmission risk estimation. The framework extracts clinically significant concepts from C-CDA sections including diagnoses, discharge medications, procedures, laboratory findings, allergies, functional status, care instructions, and follow-up plans and maps them to standardized FHIR resources such as Patient, Encounter, Condition, Observation, MedicationRequest, Procedure, CarePlan, and ServiceRequest. C2FHIR-ReadmitNet employs a clinical transformer encoder for contextual text representation, a resource-aware embedding network for structured FHIR features, a temporal attention module for modelling longitudinal encounters, and a calibrated risk-classification layer for estimating 30-day readmission probability. The proposed model is comparatively evaluated against Logistic Regression, Random Forest, XGBoost, Bidirectional Long Short-Term Memory networks, ClinicalBERT, and conventional multimodal fusion models using AUROC, AUPRC, accuracy, precision, recall, F1-score, sensitivity, specificity, Brier score, and calibration error. Comparative ROC curves, precision-recall curves, calibration plots, confusion matrices, feature-importance graphs, and ablation-analysis charts are incorporated to examine predictive discrimination, reliability, interpretability, and the contribution of individual architectural components. The study further introduces a transition-of-care risk index that integrates predicted readmission probability with medication complexity, unresolved clinical concerns, follow-up urgency, comorbidity burden, and continuity-of-care indicators to support post-discharge prioritization. The experimental design is structured to determine whether C2FHIR-ReadmitNet can achieve superior predictive discrimination, calibration, and clinical interpretability relative to conventional machine-learning and transformer-based baselines while preserving semantic interoperability across heterogeneous electronic health-record environments. The proposed approach provides a technically scalable foundation for converting C-CDA discharge information into FHIR-compatible, AI-assisted decision intelligence capable of supporting early identification of high-risk patients, targeted transitional-care interventions, interoperable clinical workflows, and data-driven reduction of avoidable hospital readmissions.