Evaluating interoperability and data quality in FHIR-based AI assessment pipelines

Abstract

We present a comprehensive implementation and evaluation of a Fast Healthcare Interoperability Resources (FHIR)–based pipeline for patient-facing AI assessment. In this pipeline, patient-reported symptoms are ingested via a FHIR-compliant REST API as Observation resources, processed by an AI inference engine, and returned as structured FHIR output (e.g. Condition or DiagnosticReport resources). We performed a synthetic comparative study against a traditional, non-standardized data exchange approach (such as ad-hoc JSON or HL7 v2), measuring key metrics: data transmission latency, information completeness, and semantic integrity. Our results show that the FHIR pipeline yields substantially higher data completeness and fidelity (capturing nearly all required fields with correct coding) compared to the legacy format, at the cost of only modest increases in payload size and processing time. In numbers, the FHIR approach retained about 95% of required data fields versus ~70% for the custom pipeline, illustrating the benefit of standardized resource profiles. These findings align with prior work on FHIR-enabled data harmonization pipelines. We conclude that using FHIR standards significantly enhances data quality and interoperability for AI-driven patient assessment, providing a reusable blueprint for clinical AI system developers. All code for pipeline diagrams and performance charts (using Graphviz, Mermaid, Matplotlib, etc.) is made available to support reproducibility.