This article addresses the problem of estimating the demand for public transport from two approaches. First, we propose a bilevel optimization problem that allows estimating the demand using historical data and the observed bus frequencies. This model has been applied to small theoretical networks and the transit network of Tandil (a medium-sized city in Buenos Aires, Argentina), showing good results. However, from a practical point of view, the computation time of the algorithm used to solve the bilevel problem is long, reducing its applicability by traffic authorities. To solve this, we propose to use an artificial neural network module that allows to quickly detect if the change in demand is significant enough (for example, beyond a predefined threshold). If it is substantial, the operator can decide to run the algorithm to estimate the demand and take action to adapt the system to the new reality, for example, adapting vehicle frequencies or incorporating more vehicles into the system so that the current demand can be served. The machine learning approach allows it to be used as a fast change detection tool, avoiding running the expensive algorithm for false positives.