FHIR-Former: enhancing clinical predictions through Fast Healthcare Interoperability Resources and large language...

Area of Science:

• Healthcare Informatics
• Artificial Intelligence in Medicine
• Clinical Data Science

Background:

• Clinical predictive modeling faces challenges with data heterogeneity and manual feature engineering.
• Existing frameworks often struggle with diverse data sources and require extensive preprocessing.
• There is a need for automated and standardized solutions for clinical prediction tasks.

Purpose of the Study:

• To introduce FHIR-Former, an open-source framework for automating and standardizing clinical prediction.
• To integrate Fast Healthcare Interoperability Resources (FHIR) with large language models (LLMs) for enhanced predictive modeling.
• To address data heterogeneity and manual feature engineering challenges in clinical settings.

Main Methods:

• FHIR-Former processes structured and unstructured data from FHIR resources dynamically.
• The framework supports multiple classification tasks, including readmission, imaging study, and ICD code prediction.
• Models were trained on 1.1 million data points using open-source LLMs (GeBERTa) and optimized via Bayesian methods.

Main Results:

• Achieved 70.7% F1-score and 72.9% accuracy for 30-day readmission prediction.
• Demonstrated 51.8% F1-score and 88.1% accuracy for mortality prediction.
• Attained 61% macro F1-score for imaging study classification and 94% accuracy for ICD code prediction.

Conclusions:

• FHIR-Former eliminates institution-specific preprocessing, adapting to diverse FHIR implementations for seamless multimodal data integration.
• The framework's configurable architecture outperforms prior methods reliant on static or limited text inputs.
• By harmonizing FHIR standardization with LLM flexibility, FHIR-Former advances scalable, interoperable predictive modeling in healthcare.