Skin disorders affect nearly 900 million individuals worldwide, but diagnostic error alone is responsible for nearly 20% of misdiagnosis and requires more accurate predictive systems to be in place. While artificial intelligence (AI) algorithms hold great potential for dermatalogic diagnosis, they are often undermined by erroneous, skewed, or wrongly labeled training datasets. Here, we introduce a selective erasure model that erases biased or corrupted training samples without full retraining. On a densely annotated dermatology image dataset of over 25,000 labeled images, our scheme enhanced diagnostic accuracy by 12.8%, reducing error rates substantially from conventional deep learning models. The method also minimized false positives by 21.4%, improving both clinical safety and credibility. Beyond performance enhancements, the methodology enables machine unlearning techniques that ensure patient anonymity, ethical application of AI, and adaptability, forming the foundation for secure and equitable skin disease prediction models.