In recent years, the incidence of Parkinson’s disease has increased linearly, and evidence-based treatment is the basic method of Parkinson’s disease Chinese medicine diagnosis and treatment. Based on Bayesian theory, this paper proposes an algorithm for the classification of TCM symptoms. Seizing the defects of uneven identification of signs and complex relations between signs, the algorithm is optimized by using the method of theme modeling, and finally an improved version of the TCM sign classification algorithm is formed. Comparing the ROC curves with DenseNet121 and DAMNet models, the improved TCM evidence classification algorithm achieves an accuracy of 96.73% and a precision of 97.45% with a minimum parameter of 7.8M. It proves that the improved TCM evidence classification algorithm is more efficient and precise in the diagnosis of Parkinson’s disease, and provides basic data and directional guidance for the future clinical application of the TCM evidence classification algorithm in the diagnosis and treatment of Parkinson’s disease.