Experimental Evaluation of Machine Learning Models for Complete Blood Count (CBC)-Based Binary Anaemia Classification

Abstract

Anaemia is a common public health problem that can benefit from accurate and timely identification to inform effective medical practice. This paper experimentally investigates the performance of machine learning models for the binary classification of anaemia based on complete blood count (CBC) information and relevant haematological variables. We created a comprehensive dataset of 2,785 CBC results obtained from publicly available sources, including haemoglobin, red blood cell count, packed cell volume, and red cell indices. After data preprocessing and correlation analysis for missing values and multicollinearity, we used principal component analysis for dimensionality reduction. We trained Logistic Regression, Random Forest, and XGBoost models on an 80:20 split for model training and testing and evaluated them using stratified 5-fold cross-validation. Accuracy, precision, recall, F1-score, and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) were used as performance metrics. The findings demonstrate that ensemble models perform better than Logistic Regression, with XGBoost performing best: accuracy of 98.92%, F1-score of 0.9906, and ROC-AUC of 1.00. The high recall of the models indicates their excellent potential for anaemia detection, while their precision indicates their suitability for accurate decision-making in anaemia diagnosis.