Machine Learning for Predicting Post-Partum Anemia: A Comparative Performance Analysis

Abstract

Anemia is a public health concern, particularly affecting post-partum women due to the increased iron demands of the fetus. Machine learning (ML) has emerged as a promising approach to address this issue by enabling early detection. However, some obstacles limit the application of ML in clinical settings. One major challenge is the inherent characteristics of the datasets used in analyses that can significantly impact the predictive performance of ML methods. Selecting the most suitable ML algorithm and considering factors such as dataset characteristics, clinical context, and intended application are crucial tasks for achieving optimal ML performance in the medical field. Our study provides a comparative performance analysis of four ML approaches - Ensemble, Rule-based, Linear, and non-linear - to select the best-performing models for the early detection of post-partum anemia, enabling timely intervention. This study was carried out in an experimental setting using a dataset comprising 355 post-partum women. The mother's age at delivery, complete blood count, and additional iron blood tests, including iron, transferrin, and ferritin, are used as features for ML. Our findings show that the Treebag algorithm outperforms all the other assessed algorithms, achieving an AUC-ROC score of approximately 0.8 in predicting post-partum anemia. Further performance metrics - F1-score (0.76), Kappa (0.6), Precision (0.72), and Recall (0.8) - confirmed that the Treebag model holds significant potential for predicting post-partum anemia and thus deserves further investigation.