Following the emergence of Elastic Optical Networks (EONs), Machine Learning (ML) has been intensively investigated as a promising methodology to address complex network management tasks, including, e.g., Quality of Transmission (QoT) estimation, fault management, and automatic adjustment of transmission parameters. Though several ML-based solutions for specific tasks have been proposed, how to integrate the outcome of such ML approaches inside Routing and Spectrum Assignment (RSA) models (which address the fundamental planning problem in EONs) is still an open research problem. In this study, we propose a dual-stage iterative RSA optimization framework that incorporates the QoT estimations provided by a ML regressor, used to define lightpaths’ reach constraints, into a Mixed Integer Linear Programming (MILP) formulation. The first stage minimizes the overall spectrum occupation, whereas the second stage maximizes the minimum inter-channel spacing between neighbor channels, without increasing the overall spectrum occupation obtained in the previous stage. During the second stage, additional interference constraints are generated, and these constraints are then added to the MILP at the next iteration round to exclude those lightpaths combinations that would exhibit unacceptable QoT. Our illustrative numerical results on realistic EON instances show that the proposed ML-assisted framework achieves spectrum occupation savings up to 52.4% (around 33% on average) in comparison to a traditional MILP-based RSA framework that uses conservative reach constraints based on margined analytical models.