Boruta is a feature selection algorithm based on random forests， whose core principle involves introducing shadow features to evaluate the importance of original features， thereby obtaining a subset of features with higher importance than their shadow counterparts. However， Boruta’s shadow feature generation mechanism involves randomly shuffling the original features， exhibiting significant randomness and uncertainty. Consequently， focusing on the shadow feature generation mechanism， a regularised autoencoder perturbation feature selection algorithm （RAE-PFS） injected with Gaussian noise was proposed. The proposed algorithm learned structural information from original features via a regularised autoencoder， encoded the original features into the latent space latent space， introduced Gaussian noise to perturb the latent space， and decoded them to generate shadow features. After merging with original features， a feature subset with higher importance than shadow features was obtained using a random forest. Experiments across eight public datasets and five classifiers demonstrate that the proposed RAE-PFS reduces the feature selection rate by 8.34% compared to traditional Boruta， while significantly improving classification accuracy， precision， and recall. Finally， interpretability analysis based on SHAP was performed on medical datasets to verify the rationality of the features selected by RAE-PFS.