Automated prediction of low ferritin concentrations using a machine learning algorithm

Area of Science:

• Biomedical Informatics
• Clinical Pathology
• Machine Learning in Healthcare

Background:

• Laboratory test interpretation can be enhanced by computational algorithms.
• Anemia is a common condition where low body iron storage (indicated by low ferritin) is a key concern.
• Primary care physicians require efficient tools to assess iron status in anemic patients.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for predicting low ferritin plasma levels.
• To utilize a minimal set of routine laboratory tests (complete blood count and C-reactive protein) for risk assessment.
• To compare the algorithm's performance against human expert interpretation.

Main Methods:

• A machine learning algorithm was developed and validated using laboratory measurements from anemic primary care patients.
• The algorithm's predictive accuracy for low ferritin was compared to twelve laboratory medicine specialists.
• Specialists' performance was assessed with and without the algorithm's output as a decision support tool.

Main Results:

• Machine learning algorithms achieved high accuracy in predicting low ferritin levels (AUC 0.92 and 0.90) on different analyzers.
• Specialists demonstrated lower accuracy than the algorithms, even when provided with the algorithm's predictions.
• Algorithm implementation led to an average of one new iron deficiency diagnosis per day.

Conclusions:

• Machine learning algorithms can accurately predict low ferritin levels in anemic patients using standard laboratory results.
• These algorithms serve as effective decision support tools, reducing unrecognized iron deficiencies.
• Integrating such algorithms into laboratory systems can improve the diagnosis and management of iron deficiency.