N-methyl-D-aspartate (NMDA) receptors are critical for cortical function and their dysfunctions are implicated in psychiatric disorders. Previous modeling studies focused on their role in generating long-timescale neuronal dynamics for working memory or evidence accumulation. Less is known about how NMDA-mediated synapses shape oscillatory dynamics in local field potentials (LFPs) or non-invasive E/MEG recordings, and how changes in NMDA transmission can be inferred from such measures of mass activity. To address these issues, we developed a recurrent network model of excitatory (E) and inhibitory (I) neurons connected by NMDA, AMPA, and GABA synapses. We computed LFP from these networks using previously validated approximations. We found that changes in the relative impact of NMDA-mediated synapses onto E vs I neurons modulate gamma-band oscillations and slope of the aperiodic component of LFP spectra (Fig 1A). Increasing NMDA relative to AMPA recurrent transmission, enhances the network-level encoding of temporal information about stimulus timescales (Fig 1B). To invert the above model, we trained a machine learning algorithm on network simulations to learn how LFP spectra depend on NMDA and then we used it to estimate NMDA changes from E/MEG data recorded from human volunteers administered either with placebo or NMDA receptor antagonists (memantine or ketamine). The algorithm predicted that both drugs decreased more NMDA-mediated transmission onto I neurons compared to E neurons, resulting in a net disinhibition (Fig 1C). Our results shed new light on how NMDA shapes network dynamics, and our approach helps to infer changes of NMDA function from E/MEG recordings.