This study assesses the connectivity alterations caused by Alzheimer's disease (AD) and mild cognitive impairment (MCI) in magnetoencephalogram (MEG) background activity. Moreover, a novel methodology to adaptively extract brain rhythms from the MEG is introduced. This methodology relies on the ability of empirical mode decomposition to isolate local signal oscillations and constrained blind source separation to extract the activity that jointly represents a subset of channels. Inter-regional MEG connectivity was analysed for 36 AD, 18 MCI and 26 control subjects in δ, θ, α and β bands over left and right central, anterior, lateral and posterior regions with magnitude squared coherence-c(f). For the sake of comparison, c(f) was calculated from the original MEG channels and from the adaptively extracted rhythms. The results indicated that AD and MCI cause slight alterations in the MEG connectivity. Computed from the extracted rhythms, c(f) distinguished AD and MCI subjects from controls with 69.4% and 77.3% accuracies, respectively, in a full leave-one-out cross-validation evaluation. These values were higher than those obtained without the proposed extraction methodology.