Abstract: 【Purpose/significance】Achieving automatic classification of short texts in medical Q&A communities is crucial for improv⁃ ing service efficiency and enhancing user experience. This paper proposes a short text classification method that combines keyword ex⁃ pansion techniques and deep learning models to address the problems of feature sparsity and semantic ambiguity in short text classifi⁃ cation.【Method/process】First, web crawlers are used to collect short user question texts from the medical Q&A community "www. xywy.com". Then, TF-IWF is applied to weight keyword importance, and FastText is used to calculate keyword similarity to expand the short text features. Next, prompt learning is integrated with deep learning models to construct a Prompt-BERT-RCNN model for ef⁃ fective classification of medical short texts.【Result/conclusion】Empirical research shows that the classification performance signifi⁃ cantly improves after keyword expansion, with the Prompt-BERT-RCNN model achieving a classification accuracy of 97.92% on the expanded medical short texts and performing excellently across nine different medical categories.【Innovation/limitation】The TFIWF and FastText-based short text expansion method solves the shortcomings of Word2vec, which does not account for keyword rarity and subword contextual information. The Prompt-BERT-RCNN model further improves classification accuracy by combining the guid⁃ ance of prompts, BERT's deep semantic understanding, and RCNN's region-aware feature extraction capabilities. However, the model's accuracy for some topics still needs improvement.