We study the minimum \emph{Monitoring Edge Geodetic Set} (\megset) problem introduced in [Foucaud et al., CALDAM'23]: given a graph $G$, we say that an edge is monitored by a pair $u,v$ of vertices if \emph{all} shortest paths between $u$ and $v$ traverse $e$; the goal of the problem consists in finding a subset $M$ of vertices of $G$ such that each edge of $G$ is monitored by at least one pair of vertices in $M$, and $|M|$ is minimized. In this paper, we prove that all polynomial-time approximation algorithms for the minimum \megset problem must have an approximation ratio of $Ω(\log n)$, unless \p = \np. To the best of our knowledge, this is the first non-constant inapproximability result known for this problem. We also strengthen the known \np-hardness of the problem on $2$-apex graphs by showing that the same result holds for $1$-apex graphs. This leaves open the problem of determining whether the problem remains \np-hard on planar (i.e., $0$-apex) graphs. On the positive side, we design an algorithm that computes good approximate solutions for hereditary graph classes that admit efficiently computable balanced separators of truly sublinear size. This immediately results in polynomial-time approximation algorithms achieving an approximation ratio of $O(n^{\frac{1}{4}} \sqrt{\log n})$ on planar graphs, graphs with bounded genus, and $k$-apex graphs with $k=O(n^{\frac{1}{4}})$. On graphs with bounded treewidth, we obtain an approximation ratio of $O(\log^{3/2} n)$ for any constant $\varepsilon > 0$. This compares favorably with the best-known approximation algorithm for general graphs, which achieves an approximation ratio of $O(\sqrt{n \log n})$ via a simple reduction to the \textsc{Set Cover} problem.