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

Background and objectives: This study aims to develop a novel platform combining machine learning and microscope images for personalized assessment of red blood cell (RBC) storage lesions. RBCs undergo storage lesions, which adversely affect transfusion outcomes. Currently, there is no individualized assessment method for RBC aging applicable in clinical practice.

Materials and methods: Blood smears and cytospin preparations from stored donor RBCs were digitized using whole-slide scanning. Predictive models were developed and validated using classical machine learning, deep learning and ensemble learning techniques. These models were tested against various datasets and validated with flow cytometry. The training dataset comprised 550,870 images, the internal testing set included 192,562 images and the external testing set contained 350,793 images. Models such as k-nearest neighbour, support vector machine, extra trees, DenseNet-121, InceptionV3 and ResNet101 were employed, with ensemble learning leveraging InceptionV3 for enhanced performance.

Results: Classical machine learning models showed modest performance, whereas deep learning models (DenseNet-121, InceptionV3, ResNet101) significantly outperformed them, achieving accuracy rates up to 0.86 on the internal testing set and 0.83 on the external testing set. The RBC morphology ensemble learning model (RBC-MELM) further enhanced predictive capabilities, particularly in the blood smear and cytospin datasets. Comparative analyses with flow cytometry indicated that while flow cytometry detected accelerated aging under certain conditions, our machine learning approaches more effectively identified RBCs exhibiting accelerated aging.

Conclusion: The proposed method utilizing machine learning techniques and microscopic blood smear analysis provides a rapid, accurate and stable approach for the personalized assessment of RBC storage lesions.