Machine Learning for Detecting Iron Deficiency through Comprehensive Blood Analysis.

Area of Science:

• Computational biology
• Medical informatics
• Hematology

Background:

• Iron deficiency (ID) is a widespread health concern.
• Current diagnostic methods for ID are limited for large-scale screening.
• Early detection of ID is essential for patient well-being.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for detecting iron deficiency (ID) in the general population.
• To utilize complete blood count (CBC) and cell population data (CPD) for ID screening.
• To enable efficient and accessible ID detection without biochemical tests.

Main Methods:

• Retrospective collection of patient data from three hospitals.
• Development and validation of five ML models using CBC, CPD, and demographic data.
• Evaluation of feature sets and subgroup performance for model robustness.

Main Results:

• The best-performing ML model achieved an AUROC > 0.94 and AUPRC > 0.83 during validation.
• Real-world deployment showed sustained performance with AUROC of 0.948 and AUPRC of 0.854.
• Model performance was lower in male and nonanemic subgroups.

Conclusions:

• An ML model integrating CPD with CBC parameters is effective for general population ID screening.
• The model facilitates efficient and consistent ID screening using routine blood data.
• This approach bypasses the need for specialized biochemical tests for ID detection.