Binary Classification vs. Anomaly Detection on Imbalanced Tabular Medical Datasets

When working on imbalanced medical datasets for a classification task, the chosen method depends mainly on label availability. We consider two classes, e.g., normal and abnormal. If the dataset is labeled, we can choose Binary Classification; otherwise, we should opt for Anomaly Detection. We aim to determine the trade-off between using labeled data with Binary Classification and unlabeled data with Anomaly Detection while being aware of the labeling effort. The comparison is made on BabyGarches, an original dataset of newborn motricity, and on two publicly available imbalanced tabular medical datasets, Thyroid and Cardioctography. We observe that Binary Classification outperforms Anomaly Detection on all datasets, achieving the highest Area Under the Curve ROC scores. Nevertheless, the performance of Deep Anomaly Detection methods such as Autoencoders and GANs is close to Binary Classification models' performance. Moreover, they reach the best Recall scores, which physicians appreciate. They may thus be recommended when data labeling is costly. Since labeling our use case dataset BabyGarches is expensive, Anomaly Detection is relevant to its classification.