Abstarct: Electronic cigarettes have gained significant popularity as an alternative to traditional tobacco smoking, particularly among young people. Understanding the chemical composition of e-liquids is essential for assessing potential health risks associated with vaping. In this study, a comprehensive quantitative structure-retention relationship (QSRR) modeling approach was developed to predict the linear retention indices (LRI) of volatile and semi-volatile organic compounds (VOCs) present in e-liquids. Experimental data were obtained from a previous study that identified 253 compounds using non-target screening with gas chromatography coupled to time-of-flight mass spectrometry (GC-TOF-MS) on an HP5-MS column. A subset of 77 compounds with reliable identification levels and LRI values was selected for modeling. Molecular structures were optimized using the semi-empirical AM1 method, and molecular descxriptors were calculated with Dragon software. After removing constant and highly collinear descxriptors (>0.95 correlation), feature selection was performed using four advanced methods: Stepwise Regression (SW), Particle Swarm Optimization (PSO), Firefly Algorithm (FF), and Random Forest (RF). Predictive models were built using multiple linear regression (MLR) and multi-laxyer perceptron artificial neural networks (MLP-ANN). Models were validated using R², adjusted R², RMSE, MAE, Q², Y-scrambling test, and applicability domain assessment via Williams plots. The best performance was achieved with PSO-ANN (R² ≈ 0.95 in test set) and RF-ANN (R² ≈ 0.92 in training set), while FF-ANN and SW-ANN showed good but comparatively lower accuracy. Y-scrambling tests yielded mean R² values close to zero (0.0004 for RF/FF and 2×10⁻⁵ for PSO), confirming the absence of chance correlations. Applicability domain coverage ranged from 92.1% to 100% across models, with no significant outliers. The results demonstrate that hybrid intelligent feature selection combined with ANN provides highly accurate, stable, and generalizable QSRR models for predicting LRI values of semi-volatile compounds in e-liquids. These models enable rapid, cost-effective screening of potentially harmful components (e.g., aldehydes, acetals, nicotine-related alkaloids, and cannabinoids) without the need for extensive experimental analysis. The developed approach offers a powerful tool for quality control, safety assessment, and regulatory monitoring of electronic cigarette liquids.