Alzheimer's disease (AD) is one of the most common brain dementia characterized by gradual deterioration of cognitive function. While it has been affecting an increasing number of aging population and become a nation-wide public health crisis, the underlying mechanism remains largely unknown. To address this problem, we propose to investigate the gene co-expression network changes along AD progression. Unlike extant work that focus on cognitive normals (CNs) and AD patients, we aim to capture the network changes during the full range of disease progression, from CN, early mild cognitive impairment (EMCI) to late MCI (LMCI) and AD. In addition, many existing differential co-expression network analyses estimate the network of each group independently, which may possibly lead to suboptimal results. Assuming that the gene co-expression patterns should be largely similar in consecutive disease stages, we propose to apply a modified joint graphical lasso model to estimate the networks of multiple diagnostic groups simultaneously. The permutation results shows that JGL model is much less likely to generate false positives with the similarity constraint. By comparing the estimated gene co-expression networks of all disease stages, we identified 8 clusters showing gradual changes during the progression of AD.