As a high-quality protein source, egg plays an important role in the global dietary structure. However, the health controversy caused by their cholesterol content has continuously attracted consumers?? attention. Traditional detection methods require sample destruction, which are time-consuming and laborintensive, and thus cannot meet the needs of large-scale grading in the poultry egg industry. To achieve rapid and non-destructive detection of cholesterol content, the visible – near-infrared spectroscopy technology was employed, combined with methods such as the least squares support vector machine (LSSVM), partial least squares regression ( PLSR ), and random forest, to construct regression prediction models for egg cholesterol content. Transmission spectra of 135 fresh eggs were collected under three postures: horizontal placement ( L), tip-down ( S), and blunt-end down ( B). Combined with preprocessing methods such as multiplicative scatter correction ( MSC) and standard normal variate transformation ( SNV), the effects of three feature selection algorithms, i. e. , competitive adaptive reweighted sampling ( CARS), successive projections algorithm ( SPA), and uninformative variable elimination (UVE) on model performance were systematically compared. The results showed that the spectral collection posture significantly affected the prediction accuracy, among which the model built based on the spectra collected in the horizontal (L) posture exhibited the optimal performance. After KS data partitioning, the SNV CARS LSSVM model achieved the best performance, with the coefficient of determination (R2) of the training set reaching 0.9190 and that of the test set being 0.9174; the root mean square error of RMSEC was 60.8682mg/(100g), and the root mean square error of RMSEP was 41.0119mg/(100g). It indicated that the established spectral model can realize rapid, non-destructive and accurate prediction of egg cholesterol content, providing technical support for egg quality monitoring and consumer choices.