Multiclass medical datasets are often imbalanced that result in underperformance of deep learning methods, especially for fewer sample classes. These underrepresented classes often correspond to rare/severe diseases, thus there is an urgent need to develop simple and effective way to address it. We propose a novel data-engineered prototypical metric approach (DEnPL). Our DEnPL consists of two parts: First, an encoder is used to generate initial feature embeddings on unperturbated data that are then averaged per class to produce class-specific prototypes, then we perform data perturbation with our proposed dual-sample augmentation averaging (DSAA) strategy allowing the model to learn data variability. The class prototypes are then separated by using an inter-class prototypical loss penalization while intra-class loss minimize the distance between the same class sample. Additionally, we introduce mean logits for samples from the DSAA block to perform classification task. DEnPL demonstrates improvement over existing methods evaluated on eight different medical datasets and shows generalizability to unseen (not used in training) datasets. Our code is available https://github.com/WANG-ZIHENG/DEnPL .