Medical multi-label classification is hampered by two intertwined data pathologies: a surfeit of “clean” images that contain no positive findings and extreme label imbalance in which clinically critical diseases appear only a handful of times. Together, these factors bias conventional CNNs toward a conservative decision boundary that overwhelmingly predicts no finding, leaving rare conditions severely under-represented in the learned feature space. We propose Counterfactual Variational autoencoder Augmentation (COVA), a lightweight pipeline that trains a variational auto-encoder on the full training set and then counterfactually edits latent codes of negative or head-class images to synthesize realistic examples of tail diseases. Each generated sample carries ground-truth labels by construction and is re-inserted into the training pool without modifying the backbone architecture or loss function. Evaluated on a de-identified chest-X-ray benchmark with 112,120 images and 14 disease labels (tail-to-head ratio \(\approx 1{:}98\) ), COVA lifts the average area under receiver-operating-characteristic curve(AUROC) from 0.799 to 0.840 \((+4.1\,\%p)\) over a strong baseline that combines a randomized augmentation policy (RandAugment) with class-weighted loss. To our knowledge, this is the first study to exploit counterfactual image synthesis for data augmentation in medical multi-label classification. These findings show that label-aware generative augmentation can meaningfully reduce false negatives and improve reliability in medical multi-label settings without additional annotation cost.