Using Machine Learning Models to Detect Anemia and Evaluate Accuracy Across Patient Subgroups in Blood Test Data

Abstract

Anemia is a widespread blood disorder currently affecting 2 billion people around the world. While machine learning (ML) has shown potential for enhanced detection of the condition using complete blood count (CBC) data, previous studies have found that ML models are unable to make consistent predictions across different demographic subgroups and patient populations. The study discussed in this paper trained and assessed two supervised ML models, Random Forest and Logistic Regression, on a CBC anemia dataset from the Eureka Diagnostic Center (EDC) in Lucknow, India, consisting of 364 patients. Hemoglobin (HGB) and features mathematically derived from it were excluded as an effort to prevent data leakage. During the study, Random Forest performed better than Logistic Regression in every accuracy metric considered for both overall testing and subgroup analysis, but both models performed poorly for elderly patients, aged 65 years and older. An ML framework, SHapley Additive exPlanations (SHAP), was used to analyze the importance of each of the features in the dataset, with red blood cell (RBC) count, packed cell volume (PCV), mean cell volume (MCV), and red cell distribution width (RDW) as the most important factors, while age and sex did not have substantial influence on the results. Both models were externally validated on a separate dataset from the Al-Zahraa Al-Ahly Hospital (AAH) in Iraq, which resulted in high accuracy and precision, but recall declined severely, suggesting that many true anemia cases were misdiagnosed when these ML models were applied to an unknown and different patient population. The results suggest that ML models could be used to detect anemia accurately within the population they were trained on, but for real-world use, the models will need to be trained on more diverse datasets.