Abstract: Based on the health-related texts from WeChat public accounts, the study conducts feature mining to explore how to effectively improve the accuracy of false health information detection, enhance the quality of healthy information on online platforms, and provide reference for users to make health-related decisions.Starting with four dimensions—content features, emotional features, publisher features and domain features, this paper proposes a health misinformation recognition method based on multidimensional feature fusion and multi-layer perceptron method(MDFF-MLP). Firstly, the textual content is analyzed to extract multidimensional key features. Secondly, feature ablation experiments are conducted to determine feature combinations with rich meaning and strong ability to distinguish true and false information. The sentiment lexicon in the dietary health domain is optimized, and a multilayer perceptron model based on the multi-dimensional fusion of “content-emotion-publisher-domain” features is constructed. Finally, a deep belief network is used to learn multidimensional features and classify dietary health misinformation. The findings indicate that in the MDFF-MLP model, the discriminative power of single-dimensional features reaches 0.90 for content features and 0.84 for publisher features. After multidimensional feature fusion, the model achieves an accuracy of 0.96, with an F1 score of 0.95. Compared with baseline models such as Logistic regression(F1=0.82) and LightGBM(F1=0.89), this model demonstrates significant improvement and strong practical applicability. 

Key words: Feature fusion, Health misinformation, Information governance, Multi-layer , perceptron, Deep learning