A roadmap to artificial intelligence (AI): Methods for designing and building AI ready data to promote fairness

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Electronic health record (EHR) data often contains biases that can negatively impact artificial intelligence (AI) model performance.
• Preparing EHR data is crucial for developing reliable AI tools in healthcare.

Purpose of the Study:

• To evaluate and create methods for preparing EHR data to minimize bias before applying AI.
• To develop a data framework for machine learning and natural language processing in predicting falls and fractures.

Main Methods:

• Incorporated strategies for multi-racial data inclusion, mixed data sources (outpatient, inpatient, structured, unstructured), and addressed missing data.
• Curated raw data using validated definitions for variables like age, race, gender, and healthcare utilization, involving clinical, statistical, and data expertise.
• Utilized machine learning for fall prediction from radiology reports and natural language processing for fracture risk assessment from DXA scan reports.

Main Results:

• Machine learning processing of over 5.3 million reports for fall prediction resulted in improved data representation and reduced missingness.
• Natural language processing algorithms achieved 98% accuracy in identifying fracture risk indicators from DXA reports.
• The developed data preparation methods are reproducible and applicable to other AI studies.

Conclusions:

• Optimal preparation of input data is essential to reduce algorithmic bias and prevent harmful AI outputs.
• Building AI-ready data frameworks enhances efficiency, transparency, and reproducibility in AI applications.
• This study emphasizes critical data curation aspects for AI implementation to mitigate bias.