Machine learning-enhanced digital microscopy for personalized assessment of red blood cell storage lesions

Area of Science:

• Hematology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Red blood cells (RBCs) develop storage lesions during preservation, negatively impacting transfusion efficacy.
• Current clinical practices lack individualized methods to assess RBC aging and storage quality.

Purpose of the Study:

• To develop a novel platform integrating machine learning (ML) and microscopy for personalized RBC storage lesion assessment.
• To create an accurate and clinically applicable method for evaluating RBC aging.

Main Methods:

• Digitization of blood smears and cytospin preparations from stored RBCs using whole-slide imaging.
• Development and validation of predictive models using classical ML, deep learning (DL), and ensemble learning.
• Testing models on large datasets (over 1 million images) and validating against flow cytometry.

Main Results:

• Deep learning models significantly outperformed classical ML, achieving high accuracy (up to 0.86).
• An ensemble learning model (RBC-MELM) demonstrated enhanced predictive power for RBC morphology.
• ML approaches proved more effective than flow cytometry in identifying accelerated RBC aging.

Conclusions:

• The developed ML platform offers a rapid, accurate, and stable method for personalized RBC storage lesion assessment.
• This approach has the potential to improve transfusion outcomes by enabling individualized quality control of RBCs.