Artificial neural networks (ANN), one of the popular machine learning techniques, have recently found various applications in the earthquake engineering field. This study compares the performance of classical (parametric) and ANN-based (non-parametric) regression models in developing efficient fragility functions for reinforced concrete (RC) frames. For this purpose, five popular intensity measures (IMs), namely peak ground acceleration (PGA), peak ground velocity (PGV), spectral acceleration (Sa) and spectral velocity (Sv) at the fundamental time period (T1) of the frame and average spectral acceleration (Saavg) are taken in the present study. The seismic performance of the mid-rise RC frames is assessed incorporating the soil-structure interaction (SSI) effects considering two soil classes (C and D). The SSI system is modelled by adopting a direct one-step approach, incorporating the soil and structure nonlinear behaviour. Further, seventy-one unscaled time histories are selected, and nonlinear time history analyses (NL-THA) are performed for the developed SSI models. In this study, peak inter-story drift ratio (PISDR) and peak roof displacement (PRD) are taken as the engineering demand parameter (EDP) to quantify the seismic response of the considered frames. The efficiency analysis results from the classical and ANN-based regression methods are compared, and it is observed that the ANN model gives lower dispersion values than the classical method. This could be because classical methods are limited to linear relationships only, whereas nonlinear relationships are considered through transfer functions in ANN models. Moreover, an attempt is made to develop an ANN-based single fragility model considering both soil classes.