This paper presents a novel approach to medical concept mapping within healthcare AI systems, focusing on improving the accuracy and efficiency of Natural Language Processing (NLP) tasks. The proposed framework integrates three levels of NLP analysis: syntax, semantics, and pragmatics, to address the challenge of mapping non-standard medical terms to expert-defined standard terms. The Byte Pair Encoding (BPE) algorithm is utilized at the syntax level to extract frequent subwords from Chinese medical terms, while Forward and Backward Maximum Matching algorithms are employed to identify subwords within input terms. A knowledge graph provides a pragmatics-level analysis by incorporating human knowledge and common sense into the system. At the semantics level, the word2vec Skip-gram model and cosine similarity are used to measure and map input terms to standard concepts. The proposed model achieves an accuracy of 96.81% in mapping non-standard terms to standard concepts, demonstrating its effectiveness in reducing errors and enhancing the performance of enterprise-level healthcare AI systems. This framework can be integrated into NLP engines to build robust medical AI systems, contributing to the advancement of healthcare technologies.