Damage accumulation during fatigue results in progressive stiffness degradation and alteration of dynamic characteristics in fibre-reinforced polymer composites. Employing such dynamic changes is a promising approach for structural health monitoring and predicting residual life. In this study, vibration-based fatigue life prediction of unidirectional glass fibre reinforced polymer (GFRP) laminates is investigated through experimental modal analysis combined with semi-empirical and machine learning models. Five GFRP specimens were fabricated and subjected to tensile, fatigue, and modal interrupted fatigue testing. Natural frequencies of the first six vibration modes were measured at different fatigue damage levels. Existing stiffness degradation models were transformed into frequency-based prediction models using the analytical relationship between stiffness degradation and natural frequency. Semi-empirical models proposed by Mao and Wu were fitted using the experimental data and validated by predicting the fatigue life of an unseen specimen. In parallel, Support Vector Regression (SVR) and Gaussian Process Regression (GPR) models were developed using normalised modal frequencies as input features, where four specimens were used for training and one specimen for validation. Both single-mode and multi-mode prediction frameworks were developed to evaluate the effect of modal feature selection on prediction accuracy. Results indicate that higher modes exhibit greater sensitivity to fatigue damage, with Mode 4 providing the most reliable predictions. Among machine learning approaches, the single-mode GPR (GPR-S) model achieved the best predictive performance with a coefficient of determination (R2) of 0.845 and a relative error (RE) of 17.57%. The semi-empirical Mao model demonstrated superior prediction capability with significantly lower relative error (2.88%), even with the truncated fatigue data. The study highlights that modal feature sensitivity plays a more critical role than algorithm complexity and demonstrates that integrating modal analysis with physics-based and data-driven approaches provides a reliable framework for residual fatigue life prediction of composite structures.