The Application of Machine‐Learning Algorithms for Multiclass Classification of Microcytic Anemia Revealed That a Minimum Required Number of Hematological Parameters Is Enough to Achieve High Diagnostic Accuracy

ABSTRACT Introduction Thalassemia is a major global health concern, as declared by the World Health Organization (WHO). Accurate screening of heterozygotes is of paramount importance for diagnostic guidance and disease mitigation. Methods Multiclass machine‐learning (ML) models were developed to classify alpha, beta, and delta/beta‐thalassemia heterozygotes, individuals with iron deficiency microcytic anemia and healthy individuals. Complete blood count (CBC) data were derived from 1518 individuals, simultaneously measured in four different hematological analyzers Sysmex K‐1000 (Sysmex), Cell‐Dyn Sapphire (Abbott), ADVIA 2120 (Siemens), BC‐6800 (Mindray). The Recursive Feature Elimination (RFE) method was applied to investigate the optimal number and combination of hematological parameters required for adequate model training, and their importance was determined using the SHAP method. Random Forest classifier and Active Learning (AL) method were used to evaluate diagnostic accuracy and enhance the model's performance, respectively. Dimensionality reduction techniques were applied for visualization purposes. Additionally, the existence of subclusters within each class was investigated using unsupervised techniques. Results Four hematological parameters (MCH, MCV, HGB, and RDW) are sufficient to classify basic types of microcytosis, achieving an accuracy of over 80%. Adding further parameters to the training process can improve the stability and reliability of the model when applied to unseen data, but does not increase test accuracy beyond 90%. Each microcytosis cluster is discriminated by assigning different weights to selected hematological parameters, forming five distinct clusters in the 2D plane (heterozygous alpha‐, beta‐, and delta/beta‐thalassemia, iron deficiency anemia, and normal individuals). The greatest overlap between clusters occurs between alpha‐ and beta‐thalassemia. The alpha‐thalassemia cluster appears to have a more diffuse scatter on the 2D plane, but two distinct subclusters are identified within this class, characterized by the differential expression of specific parameters (PCT, MPV, P‐LCR, PDW, MCV, MCH, and RBC). The developed model can estimate the probabilities of classifying a new case into the core established clusters and, in the case of alpha‐thalassemia, the possible subcluster. Conclusion The application of ML models for the automated differential diagnosis of microcytosis, using CBC data, revealed that the diagnostic accuracy is not proportionally dependent on the use of an increasing number of hematological parameters. A developed model is proposed which can correctly classify a new case in the core clusters or subclusters of patients with microcytosis.